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 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/EquivalenceClasses.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/StringExtras.h"
58 #include "llvm/Analysis/AliasAnalysis.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/PatternMatch.h"
79 #include "llvm/IR/Type.h"
80 #include "llvm/IR/Value.h"
81 #include "llvm/IR/ValueHandle.h"
82 #include "llvm/IR/Verifier.h"
83 #include "llvm/Pass.h"
84 #include "llvm/Support/BranchProbability.h"
85 #include "llvm/Support/CommandLine.h"
86 #include "llvm/Support/Debug.h"
87 #include "llvm/Support/raw_ostream.h"
88 #include "llvm/Target/TargetLibraryInfo.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/Local.h"
96 using namespace llvm::PatternMatch;
98 static cl::opt<unsigned>
99 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
100 cl::desc("Sets the SIMD width. Zero is autoselect."));
102 static cl::opt<unsigned>
103 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
104 cl::desc("Sets the vectorization unroll count. "
105 "Zero is autoselect."));
108 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
109 cl::desc("Enable if-conversion during vectorization."));
111 /// We don't vectorize loops with a known constant trip count below this number.
112 static cl::opt<unsigned>
113 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
115 cl::desc("Don't vectorize loops with a constant "
116 "trip count that is smaller than this "
119 /// This enables versioning on the strides of symbolically striding memory
120 /// accesses in code like the following.
121 /// for (i = 0; i < N; ++i)
122 /// A[i * Stride1] += B[i * Stride2] ...
124 /// Will be roughly translated to
125 /// if (Stride1 == 1 && Stride2 == 1) {
126 /// for (i = 0; i < N; i+=4)
130 static cl::opt<bool> EnableMemAccessVersioning(
131 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
132 cl::desc("Enable symblic stride memory access versioning"));
134 /// We don't unroll loops with a known constant trip count below this number.
135 static const unsigned TinyTripCountUnrollThreshold = 128;
137 /// When performing memory disambiguation checks at runtime do not make more
138 /// than this number of comparisons.
139 static const unsigned RuntimeMemoryCheckThreshold = 8;
141 /// Maximum simd width.
142 static const unsigned MaxVectorWidth = 64;
144 static cl::opt<unsigned> ForceTargetNumScalarRegs(
145 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
146 cl::desc("A flag that overrides the target's number of scalar registers."));
148 static cl::opt<unsigned> ForceTargetNumVectorRegs(
149 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
150 cl::desc("A flag that overrides the target's number of vector registers."));
152 /// Maximum vectorization unroll count.
153 static const unsigned MaxUnrollFactor = 16;
155 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
156 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's max unroll factor for scalar "
160 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
161 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's max unroll factor for "
163 "vectorized loops."));
165 static cl::opt<unsigned> ForceTargetInstructionCost(
166 "force-target-instruction-cost", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's expected cost for "
168 "an instruction to a single constant value. Mostly "
169 "useful for getting consistent testing."));
171 static cl::opt<unsigned> SmallLoopCost(
172 "small-loop-cost", cl::init(20), cl::Hidden,
173 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
175 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
176 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
177 cl::desc("Enable the use of the block frequency analysis to access PGO "
178 "heuristics minimizing code growth in cold regions and being more "
179 "aggressive in hot regions."));
181 // Runtime unroll loops for load/store throughput.
182 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
183 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
184 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
186 /// The number of stores in a loop that are allowed to need predication.
187 static cl::opt<unsigned> NumberOfStoresToPredicate(
188 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
189 cl::desc("Max number of stores to be predicated behind an if."));
191 static cl::opt<bool> EnableIndVarRegisterHeur(
192 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
193 cl::desc("Count the induction variable only once when unrolling"));
195 static cl::opt<bool> EnableCondStoresVectorization(
196 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
197 cl::desc("Enable if predication of stores during vectorization."));
201 // Forward declarations.
202 class LoopVectorizationLegality;
203 class LoopVectorizationCostModel;
205 /// InnerLoopVectorizer vectorizes loops which contain only one basic
206 /// block to a specified vectorization factor (VF).
207 /// This class performs the widening of scalars into vectors, or multiple
208 /// scalars. This class also implements the following features:
209 /// * It inserts an epilogue loop for handling loops that don't have iteration
210 /// counts that are known to be a multiple of the vectorization factor.
211 /// * It handles the code generation for reduction variables.
212 /// * Scalarization (implementation using scalars) of un-vectorizable
214 /// InnerLoopVectorizer does not perform any vectorization-legality
215 /// checks, and relies on the caller to check for the different legality
216 /// aspects. The InnerLoopVectorizer relies on the
217 /// LoopVectorizationLegality class to provide information about the induction
218 /// and reduction variables that were found to a given vectorization factor.
219 class InnerLoopVectorizer {
221 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
222 DominatorTree *DT, const DataLayout *DL,
223 const TargetLibraryInfo *TLI, unsigned VecWidth,
224 unsigned UnrollFactor)
225 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
226 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
227 OldInduction(0), WidenMap(UnrollFactor), Legal(0) {}
229 // Perform the actual loop widening (vectorization).
230 void vectorize(LoopVectorizationLegality *L) {
232 // Create a new empty loop. Unlink the old loop and connect the new one.
234 // Widen each instruction in the old loop to a new one in the new loop.
235 // Use the Legality module to find the induction and reduction variables.
237 // Register the new loop and update the analysis passes.
241 virtual ~InnerLoopVectorizer() {}
244 /// A small list of PHINodes.
245 typedef SmallVector<PHINode*, 4> PhiVector;
246 /// When we unroll loops we have multiple vector values for each scalar.
247 /// This data structure holds the unrolled and vectorized values that
248 /// originated from one scalar instruction.
249 typedef SmallVector<Value*, 2> VectorParts;
251 // When we if-convert we need create edge masks. We have to cache values so
252 // that we don't end up with exponential recursion/IR.
253 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
254 VectorParts> EdgeMaskCache;
256 /// \brief Add code that checks at runtime if the accessed arrays overlap.
258 /// Returns a pair of instructions where the first element is the first
259 /// instruction generated in possibly a sequence of instructions and the
260 /// second value is the final comparator value or NULL if no check is needed.
261 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
263 /// \brief Add checks for strides that where assumed to be 1.
265 /// Returns the last check instruction and the first check instruction in the
266 /// pair as (first, last).
267 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
269 /// Create an empty loop, based on the loop ranges of the old loop.
270 void createEmptyLoop();
271 /// Copy and widen the instructions from the old loop.
272 virtual void vectorizeLoop();
274 /// \brief The Loop exit block may have single value PHI nodes where the
275 /// incoming value is 'Undef'. While vectorizing we only handled real values
276 /// that were defined inside the loop. Here we fix the 'undef case'.
280 /// A helper function that computes the predicate of the block BB, assuming
281 /// that the header block of the loop is set to True. It returns the *entry*
282 /// mask for the block BB.
283 VectorParts createBlockInMask(BasicBlock *BB);
284 /// A helper function that computes the predicate of the edge between SRC
286 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
288 /// A helper function to vectorize a single BB within the innermost loop.
289 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
291 /// Vectorize a single PHINode in a block. This method handles the induction
292 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
293 /// arbitrary length vectors.
294 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
295 unsigned UF, unsigned VF, PhiVector *PV);
297 /// Insert the new loop to the loop hierarchy and pass manager
298 /// and update the analysis passes.
299 void updateAnalysis();
301 /// This instruction is un-vectorizable. Implement it as a sequence
302 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
303 /// scalarized instruction behind an if block predicated on the control
304 /// dependence of the instruction.
305 virtual void scalarizeInstruction(Instruction *Instr,
306 bool IfPredicateStore=false);
308 /// Vectorize Load and Store instructions,
309 virtual void vectorizeMemoryInstruction(Instruction *Instr);
311 /// Create a broadcast instruction. This method generates a broadcast
312 /// instruction (shuffle) for loop invariant values and for the induction
313 /// value. If this is the induction variable then we extend it to N, N+1, ...
314 /// this is needed because each iteration in the loop corresponds to a SIMD
316 virtual Value *getBroadcastInstrs(Value *V);
318 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
319 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
320 /// The sequence starts at StartIndex.
321 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
323 /// When we go over instructions in the basic block we rely on previous
324 /// values within the current basic block or on loop invariant values.
325 /// When we widen (vectorize) values we place them in the map. If the values
326 /// are not within the map, they have to be loop invariant, so we simply
327 /// broadcast them into a vector.
328 VectorParts &getVectorValue(Value *V);
330 /// Generate a shuffle sequence that will reverse the vector Vec.
331 virtual Value *reverseVector(Value *Vec);
333 /// This is a helper class that holds the vectorizer state. It maps scalar
334 /// instructions to vector instructions. When the code is 'unrolled' then
335 /// then a single scalar value is mapped to multiple vector parts. The parts
336 /// are stored in the VectorPart type.
338 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
340 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
342 /// \return True if 'Key' is saved in the Value Map.
343 bool has(Value *Key) const { return MapStorage.count(Key); }
345 /// Initializes a new entry in the map. Sets all of the vector parts to the
346 /// save value in 'Val'.
347 /// \return A reference to a vector with splat values.
348 VectorParts &splat(Value *Key, Value *Val) {
349 VectorParts &Entry = MapStorage[Key];
350 Entry.assign(UF, Val);
354 ///\return A reference to the value that is stored at 'Key'.
355 VectorParts &get(Value *Key) {
356 VectorParts &Entry = MapStorage[Key];
359 assert(Entry.size() == UF);
364 /// The unroll factor. Each entry in the map stores this number of vector
368 /// Map storage. We use std::map and not DenseMap because insertions to a
369 /// dense map invalidates its iterators.
370 std::map<Value *, VectorParts> MapStorage;
373 /// The original loop.
375 /// Scev analysis to use.
382 const DataLayout *DL;
383 /// Target Library Info.
384 const TargetLibraryInfo *TLI;
386 /// The vectorization SIMD factor to use. Each vector will have this many
391 /// The vectorization unroll factor to use. Each scalar is vectorized to this
392 /// many different vector instructions.
395 /// The builder that we use
398 // --- Vectorization state ---
400 /// The vector-loop preheader.
401 BasicBlock *LoopVectorPreHeader;
402 /// The scalar-loop preheader.
403 BasicBlock *LoopScalarPreHeader;
404 /// Middle Block between the vector and the scalar.
405 BasicBlock *LoopMiddleBlock;
406 ///The ExitBlock of the scalar loop.
407 BasicBlock *LoopExitBlock;
408 ///The vector loop body.
409 SmallVector<BasicBlock *, 4> LoopVectorBody;
410 ///The scalar loop body.
411 BasicBlock *LoopScalarBody;
412 /// A list of all bypass blocks. The first block is the entry of the loop.
413 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
415 /// The new Induction variable which was added to the new block.
417 /// The induction variable of the old basic block.
418 PHINode *OldInduction;
419 /// Holds the extended (to the widest induction type) start index.
421 /// Maps scalars to widened vectors.
423 EdgeMaskCache MaskCache;
425 LoopVectorizationLegality *Legal;
428 class InnerLoopUnroller : public InnerLoopVectorizer {
430 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
431 DominatorTree *DT, const DataLayout *DL,
432 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
433 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
436 void scalarizeInstruction(Instruction *Instr,
437 bool IfPredicateStore = false) override;
438 void vectorizeMemoryInstruction(Instruction *Instr) override;
439 Value *getBroadcastInstrs(Value *V) override;
440 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
441 Value *reverseVector(Value *Vec) override;
444 /// \brief Look for a meaningful debug location on the instruction or it's
446 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
451 if (I->getDebugLoc() != Empty)
454 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
455 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
456 if (OpInst->getDebugLoc() != Empty)
463 /// \brief Set the debug location in the builder using the debug location in the
465 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
466 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
467 B.SetCurrentDebugLocation(Inst->getDebugLoc());
469 B.SetCurrentDebugLocation(DebugLoc());
472 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
473 /// to what vectorization factor.
474 /// This class does not look at the profitability of vectorization, only the
475 /// legality. This class has two main kinds of checks:
476 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
477 /// will change the order of memory accesses in a way that will change the
478 /// correctness of the program.
479 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
480 /// checks for a number of different conditions, such as the availability of a
481 /// single induction variable, that all types are supported and vectorize-able,
482 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
483 /// This class is also used by InnerLoopVectorizer for identifying
484 /// induction variable and the different reduction variables.
485 class LoopVectorizationLegality {
489 unsigned NumPredStores;
491 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
492 DominatorTree *DT, TargetLibraryInfo *TLI)
493 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
494 DT(DT), TLI(TLI), Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
495 MaxSafeDepDistBytes(-1U) {}
497 /// This enum represents the kinds of reductions that we support.
499 RK_NoReduction, ///< Not a reduction.
500 RK_IntegerAdd, ///< Sum of integers.
501 RK_IntegerMult, ///< Product of integers.
502 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
503 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
504 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
505 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
506 RK_FloatAdd, ///< Sum of floats.
507 RK_FloatMult, ///< Product of floats.
508 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
511 /// This enum represents the kinds of inductions that we support.
513 IK_NoInduction, ///< Not an induction variable.
514 IK_IntInduction, ///< Integer induction variable. Step = 1.
515 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
516 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
517 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
520 // This enum represents the kind of minmax reduction.
521 enum MinMaxReductionKind {
531 /// This struct holds information about reduction variables.
532 struct ReductionDescriptor {
533 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
534 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
536 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
537 MinMaxReductionKind MK)
538 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
540 // The starting value of the reduction.
541 // It does not have to be zero!
542 TrackingVH<Value> StartValue;
543 // The instruction who's value is used outside the loop.
544 Instruction *LoopExitInstr;
545 // The kind of the reduction.
547 // If this a min/max reduction the kind of reduction.
548 MinMaxReductionKind MinMaxKind;
551 /// This POD struct holds information about a potential reduction operation.
552 struct ReductionInstDesc {
553 ReductionInstDesc(bool IsRedux, Instruction *I) :
554 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
556 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
557 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
559 // Is this instruction a reduction candidate.
561 // The last instruction in a min/max pattern (select of the select(icmp())
562 // pattern), or the current reduction instruction otherwise.
563 Instruction *PatternLastInst;
564 // If this is a min/max pattern the comparison predicate.
565 MinMaxReductionKind MinMaxKind;
568 /// This struct holds information about the memory runtime legality
569 /// check that a group of pointers do not overlap.
570 struct RuntimePointerCheck {
571 RuntimePointerCheck() : Need(false) {}
573 /// Reset the state of the pointer runtime information.
580 DependencySetId.clear();
583 /// Insert a pointer and calculate the start and end SCEVs.
584 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
585 unsigned DepSetId, ValueToValueMap &Strides);
587 /// This flag indicates if we need to add the runtime check.
589 /// Holds the pointers that we need to check.
590 SmallVector<TrackingVH<Value>, 2> Pointers;
591 /// Holds the pointer value at the beginning of the loop.
592 SmallVector<const SCEV*, 2> Starts;
593 /// Holds the pointer value at the end of the loop.
594 SmallVector<const SCEV*, 2> Ends;
595 /// Holds the information if this pointer is used for writing to memory.
596 SmallVector<bool, 2> IsWritePtr;
597 /// Holds the id of the set of pointers that could be dependent because of a
598 /// shared underlying object.
599 SmallVector<unsigned, 2> DependencySetId;
602 /// A struct for saving information about induction variables.
603 struct InductionInfo {
604 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
605 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
607 TrackingVH<Value> StartValue;
612 /// ReductionList contains the reduction descriptors for all
613 /// of the reductions that were found in the loop.
614 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
616 /// InductionList saves induction variables and maps them to the
617 /// induction descriptor.
618 typedef MapVector<PHINode*, InductionInfo> InductionList;
620 /// Returns true if it is legal to vectorize this loop.
621 /// This does not mean that it is profitable to vectorize this
622 /// loop, only that it is legal to do so.
625 /// Returns the Induction variable.
626 PHINode *getInduction() { return Induction; }
628 /// Returns the reduction variables found in the loop.
629 ReductionList *getReductionVars() { return &Reductions; }
631 /// Returns the induction variables found in the loop.
632 InductionList *getInductionVars() { return &Inductions; }
634 /// Returns the widest induction type.
635 Type *getWidestInductionType() { return WidestIndTy; }
637 /// Returns True if V is an induction variable in this loop.
638 bool isInductionVariable(const Value *V);
640 /// Return true if the block BB needs to be predicated in order for the loop
641 /// to be vectorized.
642 bool blockNeedsPredication(BasicBlock *BB);
644 /// Check if this pointer is consecutive when vectorizing. This happens
645 /// when the last index of the GEP is the induction variable, or that the
646 /// pointer itself is an induction variable.
647 /// This check allows us to vectorize A[idx] into a wide load/store.
649 /// 0 - Stride is unknown or non-consecutive.
650 /// 1 - Address is consecutive.
651 /// -1 - Address is consecutive, and decreasing.
652 int isConsecutivePtr(Value *Ptr);
654 /// Returns true if the value V is uniform within the loop.
655 bool isUniform(Value *V);
657 /// Returns true if this instruction will remain scalar after vectorization.
658 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
660 /// Returns the information that we collected about runtime memory check.
661 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
663 /// This function returns the identity element (or neutral element) for
665 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
667 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
669 bool hasStride(Value *V) { return StrideSet.count(V); }
670 bool mustCheckStrides() { return !StrideSet.empty(); }
671 SmallPtrSet<Value *, 8>::iterator strides_begin() {
672 return StrideSet.begin();
674 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
677 /// Check if a single basic block loop is vectorizable.
678 /// At this point we know that this is a loop with a constant trip count
679 /// and we only need to check individual instructions.
680 bool canVectorizeInstrs();
682 /// When we vectorize loops we may change the order in which
683 /// we read and write from memory. This method checks if it is
684 /// legal to vectorize the code, considering only memory constrains.
685 /// Returns true if the loop is vectorizable
686 bool canVectorizeMemory();
688 /// Return true if we can vectorize this loop using the IF-conversion
690 bool canVectorizeWithIfConvert();
692 /// Collect the variables that need to stay uniform after vectorization.
693 void collectLoopUniforms();
695 /// Return true if all of the instructions in the block can be speculatively
696 /// executed. \p SafePtrs is a list of addresses that are known to be legal
697 /// and we know that we can read from them without segfault.
698 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
700 /// Returns True, if 'Phi' is the kind of reduction variable for type
701 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
702 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
703 /// Returns a struct describing if the instruction 'I' can be a reduction
704 /// variable of type 'Kind'. If the reduction is a min/max pattern of
705 /// select(icmp()) this function advances the instruction pointer 'I' from the
706 /// compare instruction to the select instruction and stores this pointer in
707 /// 'PatternLastInst' member of the returned struct.
708 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
709 ReductionInstDesc &Desc);
710 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
711 /// pattern corresponding to a min(X, Y) or max(X, Y).
712 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
713 ReductionInstDesc &Prev);
714 /// Returns the induction kind of Phi. This function may return NoInduction
715 /// if the PHI is not an induction variable.
716 InductionKind isInductionVariable(PHINode *Phi);
718 /// \brief Collect memory access with loop invariant strides.
720 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
722 void collectStridedAcccess(Value *LoadOrStoreInst);
724 /// The loop that we evaluate.
728 /// DataLayout analysis.
729 const DataLayout *DL;
732 /// Target Library Info.
733 TargetLibraryInfo *TLI;
735 // --- vectorization state --- //
737 /// Holds the integer induction variable. This is the counter of the
740 /// Holds the reduction variables.
741 ReductionList Reductions;
742 /// Holds all of the induction variables that we found in the loop.
743 /// Notice that inductions don't need to start at zero and that induction
744 /// variables can be pointers.
745 InductionList Inductions;
746 /// Holds the widest induction type encountered.
749 /// Allowed outside users. This holds the reduction
750 /// vars which can be accessed from outside the loop.
751 SmallPtrSet<Value*, 4> AllowedExit;
752 /// This set holds the variables which are known to be uniform after
754 SmallPtrSet<Instruction*, 4> Uniforms;
755 /// We need to check that all of the pointers in this list are disjoint
757 RuntimePointerCheck PtrRtCheck;
758 /// Can we assume the absence of NaNs.
759 bool HasFunNoNaNAttr;
761 unsigned MaxSafeDepDistBytes;
763 ValueToValueMap Strides;
764 SmallPtrSet<Value *, 8> StrideSet;
767 /// LoopVectorizationCostModel - estimates the expected speedups due to
769 /// In many cases vectorization is not profitable. This can happen because of
770 /// a number of reasons. In this class we mainly attempt to predict the
771 /// expected speedup/slowdowns due to the supported instruction set. We use the
772 /// TargetTransformInfo to query the different backends for the cost of
773 /// different operations.
774 class LoopVectorizationCostModel {
776 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
777 LoopVectorizationLegality *Legal,
778 const TargetTransformInfo &TTI,
779 const DataLayout *DL, const TargetLibraryInfo *TLI)
780 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
782 /// Information about vectorization costs
783 struct VectorizationFactor {
784 unsigned Width; // Vector width with best cost
785 unsigned Cost; // Cost of the loop with that width
787 /// \return The most profitable vectorization factor and the cost of that VF.
788 /// This method checks every power of two up to VF. If UserVF is not ZERO
789 /// then this vectorization factor will be selected if vectorization is
791 VectorizationFactor selectVectorizationFactor(bool OptForSize,
794 /// \return The size (in bits) of the widest type in the code that
795 /// needs to be vectorized. We ignore values that remain scalar such as
796 /// 64 bit loop indices.
797 unsigned getWidestType();
799 /// \return The most profitable unroll factor.
800 /// If UserUF is non-zero then this method finds the best unroll-factor
801 /// based on register pressure and other parameters.
802 /// VF and LoopCost are the selected vectorization factor and the cost of the
804 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
807 /// \brief A struct that represents some properties of the register usage
809 struct RegisterUsage {
810 /// Holds the number of loop invariant values that are used in the loop.
811 unsigned LoopInvariantRegs;
812 /// Holds the maximum number of concurrent live intervals in the loop.
813 unsigned MaxLocalUsers;
814 /// Holds the number of instructions in the loop.
815 unsigned NumInstructions;
818 /// \return information about the register usage of the loop.
819 RegisterUsage calculateRegisterUsage();
822 /// Returns the expected execution cost. The unit of the cost does
823 /// not matter because we use the 'cost' units to compare different
824 /// vector widths. The cost that is returned is *not* normalized by
825 /// the factor width.
826 unsigned expectedCost(unsigned VF);
828 /// Returns the execution time cost of an instruction for a given vector
829 /// width. Vector width of one means scalar.
830 unsigned getInstructionCost(Instruction *I, unsigned VF);
832 /// A helper function for converting Scalar types to vector types.
833 /// If the incoming type is void, we return void. If the VF is 1, we return
835 static Type* ToVectorTy(Type *Scalar, unsigned VF);
837 /// Returns whether the instruction is a load or store and will be a emitted
838 /// as a vector operation.
839 bool isConsecutiveLoadOrStore(Instruction *I);
841 /// The loop that we evaluate.
845 /// Loop Info analysis.
847 /// Vectorization legality.
848 LoopVectorizationLegality *Legal;
849 /// Vector target information.
850 const TargetTransformInfo &TTI;
851 /// Target data layout information.
852 const DataLayout *DL;
853 /// Target Library Info.
854 const TargetLibraryInfo *TLI;
857 /// Utility class for getting and setting loop vectorizer hints in the form
858 /// of loop metadata.
859 struct LoopVectorizeHints {
860 /// Vectorization width.
862 /// Vectorization unroll factor.
864 /// Vectorization forced (-1 not selected, 0 force disabled, 1 force enabled)
867 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
868 : Width(VectorizationFactor)
869 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
871 , LoopID(L->getLoopID()) {
873 // The command line options override any loop metadata except for when
874 // width == 1 which is used to indicate the loop is already vectorized.
875 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
876 Width = VectorizationFactor;
877 if (VectorizationUnroll.getNumOccurrences() > 0)
878 Unroll = VectorizationUnroll;
880 DEBUG(if (DisableUnrolling && Unroll == 1)
881 dbgs() << "LV: Unrolling disabled by the pass manager\n");
884 /// Return the loop vectorizer metadata prefix.
885 static StringRef Prefix() { return "llvm.vectorizer."; }
887 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
888 SmallVector<Value*, 2> Vals;
889 Vals.push_back(MDString::get(Context, Name));
890 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
891 return MDNode::get(Context, Vals);
894 /// Mark the loop L as already vectorized by setting the width to 1.
895 void setAlreadyVectorized(Loop *L) {
896 LLVMContext &Context = L->getHeader()->getContext();
900 // Create a new loop id with one more operand for the already_vectorized
901 // hint. If the loop already has a loop id then copy the existing operands.
902 SmallVector<Value*, 4> Vals(1);
904 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
905 Vals.push_back(LoopID->getOperand(i));
907 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
908 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
910 MDNode *NewLoopID = MDNode::get(Context, Vals);
911 // Set operand 0 to refer to the loop id itself.
912 NewLoopID->replaceOperandWith(0, NewLoopID);
914 L->setLoopID(NewLoopID);
916 LoopID->replaceAllUsesWith(NewLoopID);
924 /// Find hints specified in the loop metadata.
925 void getHints(const Loop *L) {
929 // First operand should refer to the loop id itself.
930 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
931 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
933 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
934 const MDString *S = 0;
935 SmallVector<Value*, 4> Args;
937 // The expected hint is either a MDString or a MDNode with the first
938 // operand a MDString.
939 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
940 if (!MD || MD->getNumOperands() == 0)
942 S = dyn_cast<MDString>(MD->getOperand(0));
943 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
944 Args.push_back(MD->getOperand(i));
946 S = dyn_cast<MDString>(LoopID->getOperand(i));
947 assert(Args.size() == 0 && "too many arguments for MDString");
953 // Check if the hint starts with the vectorizer prefix.
954 StringRef Hint = S->getString();
955 if (!Hint.startswith(Prefix()))
957 // Remove the prefix.
958 Hint = Hint.substr(Prefix().size(), StringRef::npos);
960 if (Args.size() == 1)
961 getHint(Hint, Args[0]);
965 // Check string hint with one operand.
966 void getHint(StringRef Hint, Value *Arg) {
967 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
969 unsigned Val = C->getZExtValue();
971 if (Hint == "width") {
972 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
975 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
976 } else if (Hint == "unroll") {
977 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
980 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
981 } else if (Hint == "enable") {
982 if (C->getBitWidth() == 1)
985 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
987 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
992 static void addInnerLoop(Loop *L, SmallVectorImpl<Loop *> &V) {
994 return V.push_back(L);
996 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
1000 /// The LoopVectorize Pass.
1001 struct LoopVectorize : public FunctionPass {
1002 /// Pass identification, replacement for typeid
1005 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1007 DisableUnrolling(NoUnrolling),
1008 AlwaysVectorize(AlwaysVectorize) {
1009 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1012 ScalarEvolution *SE;
1013 const DataLayout *DL;
1015 TargetTransformInfo *TTI;
1017 BlockFrequencyInfo *BFI;
1018 TargetLibraryInfo *TLI;
1019 bool DisableUnrolling;
1020 bool AlwaysVectorize;
1022 BlockFrequency ColdEntryFreq;
1024 bool runOnFunction(Function &F) override {
1025 SE = &getAnalysis<ScalarEvolution>();
1026 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1027 DL = DLP ? &DLP->getDataLayout() : 0;
1028 LI = &getAnalysis<LoopInfo>();
1029 TTI = &getAnalysis<TargetTransformInfo>();
1030 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1031 BFI = &getAnalysis<BlockFrequencyInfo>();
1032 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1034 // Compute some weights outside of the loop over the loops. Compute this
1035 // using a BranchProbability to re-use its scaling math.
1036 const BranchProbability ColdProb(1, 5); // 20%
1037 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1039 // If the target claims to have no vector registers don't attempt
1041 if (!TTI->getNumberOfRegisters(true))
1045 DEBUG(dbgs() << "LV: Not vectorizing: Missing data layout\n");
1049 // Build up a worklist of inner-loops to vectorize. This is necessary as
1050 // the act of vectorizing or partially unrolling a loop creates new loops
1051 // and can invalidate iterators across the loops.
1052 SmallVector<Loop *, 8> Worklist;
1054 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
1055 addInnerLoop(*I, Worklist);
1057 // Now walk the identified inner loops.
1058 bool Changed = false;
1059 while (!Worklist.empty())
1060 Changed |= processLoop(Worklist.pop_back_val());
1062 // Process each loop nest in the function.
1066 bool processLoop(Loop *L) {
1067 // We only handle inner loops, so if there are children just recurse.
1069 bool Changed = false;
1070 for (Loop::iterator I = L->begin(), E = L->begin(); I != E; ++I)
1071 Changed |= processLoop(*I);
1075 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
1076 L->getHeader()->getParent()->getName() << "\"\n");
1078 LoopVectorizeHints Hints(L, DisableUnrolling);
1080 if (Hints.Force == 0) {
1081 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1085 if (!AlwaysVectorize && Hints.Force != 1) {
1086 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1090 if (Hints.Width == 1 && Hints.Unroll == 1) {
1091 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1095 // Check if it is legal to vectorize the loop.
1096 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1097 if (!LVL.canVectorize()) {
1098 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1102 // Use the cost model.
1103 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1105 // Check the function attributes to find out if this function should be
1106 // optimized for size.
1107 Function *F = L->getHeader()->getParent();
1109 Hints.Force != 1 && F->hasFnAttribute(Attribute::OptimizeForSize);
1111 // Compute the weighted frequency of this loop being executed and see if it
1112 // is less than 20% of the function entry baseline frequency. Note that we
1113 // always have a canonical loop here because we think we *can* vectoriez.
1114 // FIXME: This is hidden behind a flag due to pervasive problems with
1115 // exactly what block frequency models.
1116 if (LoopVectorizeWithBlockFrequency) {
1117 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1118 if (Hints.Force != 1 && LoopEntryFreq < ColdEntryFreq)
1122 // Check the function attributes to see if implicit floats are allowed.a
1123 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1124 // an integer loop and the vector instructions selected are purely integer
1125 // vector instructions?
1126 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1127 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1128 "attribute is used.\n");
1132 // Select the optimal vectorization factor.
1133 LoopVectorizationCostModel::VectorizationFactor VF;
1134 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
1135 // Select the unroll factor.
1136 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1139 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
1140 F->getParent()->getModuleIdentifier() << '\n');
1141 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1143 if (VF.Width == 1) {
1144 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1147 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1148 // We decided not to vectorize, but we may want to unroll.
1149 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1150 Unroller.vectorize(&LVL);
1152 // If we decided that it is *legal* to vectorize the loop then do it.
1153 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1157 // Mark the loop as already vectorized to avoid vectorizing again.
1158 Hints.setAlreadyVectorized(L);
1160 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1164 void getAnalysisUsage(AnalysisUsage &AU) const override {
1165 AU.addRequiredID(LoopSimplifyID);
1166 AU.addRequiredID(LCSSAID);
1167 AU.addRequired<BlockFrequencyInfo>();
1168 AU.addRequired<DominatorTreeWrapperPass>();
1169 AU.addRequired<LoopInfo>();
1170 AU.addRequired<ScalarEvolution>();
1171 AU.addRequired<TargetTransformInfo>();
1172 AU.addPreserved<LoopInfo>();
1173 AU.addPreserved<DominatorTreeWrapperPass>();
1178 } // end anonymous namespace
1180 //===----------------------------------------------------------------------===//
1181 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1182 // LoopVectorizationCostModel.
1183 //===----------------------------------------------------------------------===//
1185 static Value *stripIntegerCast(Value *V) {
1186 if (CastInst *CI = dyn_cast<CastInst>(V))
1187 if (CI->getOperand(0)->getType()->isIntegerTy())
1188 return CI->getOperand(0);
1192 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1194 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1196 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1197 ValueToValueMap &PtrToStride,
1198 Value *Ptr, Value *OrigPtr = 0) {
1200 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1202 // If there is an entry in the map return the SCEV of the pointer with the
1203 // symbolic stride replaced by one.
1204 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1205 if (SI != PtrToStride.end()) {
1206 Value *StrideVal = SI->second;
1209 StrideVal = stripIntegerCast(StrideVal);
1211 // Replace symbolic stride by one.
1212 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1213 ValueToValueMap RewriteMap;
1214 RewriteMap[StrideVal] = One;
1217 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1218 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1223 // Otherwise, just return the SCEV of the original pointer.
1224 return SE->getSCEV(Ptr);
1227 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1228 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1229 ValueToValueMap &Strides) {
1230 // Get the stride replaced scev.
1231 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1232 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1233 assert(AR && "Invalid addrec expression");
1234 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1235 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1236 Pointers.push_back(Ptr);
1237 Starts.push_back(AR->getStart());
1238 Ends.push_back(ScEnd);
1239 IsWritePtr.push_back(WritePtr);
1240 DependencySetId.push_back(DepSetId);
1243 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1244 // We need to place the broadcast of invariant variables outside the loop.
1245 Instruction *Instr = dyn_cast<Instruction>(V);
1247 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1248 Instr->getParent()) != LoopVectorBody.end());
1249 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1251 // Place the code for broadcasting invariant variables in the new preheader.
1252 IRBuilder<>::InsertPointGuard Guard(Builder);
1254 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1256 // Broadcast the scalar into all locations in the vector.
1257 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1262 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1264 assert(Val->getType()->isVectorTy() && "Must be a vector");
1265 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1266 "Elem must be an integer");
1267 // Create the types.
1268 Type *ITy = Val->getType()->getScalarType();
1269 VectorType *Ty = cast<VectorType>(Val->getType());
1270 int VLen = Ty->getNumElements();
1271 SmallVector<Constant*, 8> Indices;
1273 // Create a vector of consecutive numbers from zero to VF.
1274 for (int i = 0; i < VLen; ++i) {
1275 int64_t Idx = Negate ? (-i) : i;
1276 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1279 // Add the consecutive indices to the vector value.
1280 Constant *Cv = ConstantVector::get(Indices);
1281 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1282 return Builder.CreateAdd(Val, Cv, "induction");
1285 /// \brief Find the operand of the GEP that should be checked for consecutive
1286 /// stores. This ignores trailing indices that have no effect on the final
1288 static unsigned getGEPInductionOperand(const DataLayout *DL,
1289 const GetElementPtrInst *Gep) {
1290 unsigned LastOperand = Gep->getNumOperands() - 1;
1291 unsigned GEPAllocSize = DL->getTypeAllocSize(
1292 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1294 // Walk backwards and try to peel off zeros.
1295 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1296 // Find the type we're currently indexing into.
1297 gep_type_iterator GEPTI = gep_type_begin(Gep);
1298 std::advance(GEPTI, LastOperand - 1);
1300 // If it's a type with the same allocation size as the result of the GEP we
1301 // can peel off the zero index.
1302 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1310 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1311 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1312 // Make sure that the pointer does not point to structs.
1313 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1316 // If this value is a pointer induction variable we know it is consecutive.
1317 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1318 if (Phi && Inductions.count(Phi)) {
1319 InductionInfo II = Inductions[Phi];
1320 if (IK_PtrInduction == II.IK)
1322 else if (IK_ReversePtrInduction == II.IK)
1326 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1330 unsigned NumOperands = Gep->getNumOperands();
1331 Value *GpPtr = Gep->getPointerOperand();
1332 // If this GEP value is a consecutive pointer induction variable and all of
1333 // the indices are constant then we know it is consecutive. We can
1334 Phi = dyn_cast<PHINode>(GpPtr);
1335 if (Phi && Inductions.count(Phi)) {
1337 // Make sure that the pointer does not point to structs.
1338 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1339 if (GepPtrType->getElementType()->isAggregateType())
1342 // Make sure that all of the index operands are loop invariant.
1343 for (unsigned i = 1; i < NumOperands; ++i)
1344 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1347 InductionInfo II = Inductions[Phi];
1348 if (IK_PtrInduction == II.IK)
1350 else if (IK_ReversePtrInduction == II.IK)
1354 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1356 // Check that all of the gep indices are uniform except for our induction
1358 for (unsigned i = 0; i != NumOperands; ++i)
1359 if (i != InductionOperand &&
1360 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1363 // We can emit wide load/stores only if the last non-zero index is the
1364 // induction variable.
1365 const SCEV *Last = 0;
1366 if (!Strides.count(Gep))
1367 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1369 // Because of the multiplication by a stride we can have a s/zext cast.
1370 // We are going to replace this stride by 1 so the cast is safe to ignore.
1372 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1373 // %0 = trunc i64 %indvars.iv to i32
1374 // %mul = mul i32 %0, %Stride1
1375 // %idxprom = zext i32 %mul to i64 << Safe cast.
1376 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1378 Last = replaceSymbolicStrideSCEV(SE, Strides,
1379 Gep->getOperand(InductionOperand), Gep);
1380 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1382 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1386 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1387 const SCEV *Step = AR->getStepRecurrence(*SE);
1389 // The memory is consecutive because the last index is consecutive
1390 // and all other indices are loop invariant.
1393 if (Step->isAllOnesValue())
1400 bool LoopVectorizationLegality::isUniform(Value *V) {
1401 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1404 InnerLoopVectorizer::VectorParts&
1405 InnerLoopVectorizer::getVectorValue(Value *V) {
1406 assert(V != Induction && "The new induction variable should not be used.");
1407 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1409 // If we have a stride that is replaced by one, do it here.
1410 if (Legal->hasStride(V))
1411 V = ConstantInt::get(V->getType(), 1);
1413 // If we have this scalar in the map, return it.
1414 if (WidenMap.has(V))
1415 return WidenMap.get(V);
1417 // If this scalar is unknown, assume that it is a constant or that it is
1418 // loop invariant. Broadcast V and save the value for future uses.
1419 Value *B = getBroadcastInstrs(V);
1420 return WidenMap.splat(V, B);
1423 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1424 assert(Vec->getType()->isVectorTy() && "Invalid type");
1425 SmallVector<Constant*, 8> ShuffleMask;
1426 for (unsigned i = 0; i < VF; ++i)
1427 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1429 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1430 ConstantVector::get(ShuffleMask),
1434 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1435 // Attempt to issue a wide load.
1436 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1437 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1439 assert((LI || SI) && "Invalid Load/Store instruction");
1441 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1442 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1443 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1444 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1445 // An alignment of 0 means target abi alignment. We need to use the scalar's
1446 // target abi alignment in such a case.
1448 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1449 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1450 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1451 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1453 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1454 return scalarizeInstruction(Instr, true);
1456 if (ScalarAllocatedSize != VectorElementSize)
1457 return scalarizeInstruction(Instr);
1459 // If the pointer is loop invariant or if it is non-consecutive,
1460 // scalarize the load.
1461 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1462 bool Reverse = ConsecutiveStride < 0;
1463 bool UniformLoad = LI && Legal->isUniform(Ptr);
1464 if (!ConsecutiveStride || UniformLoad)
1465 return scalarizeInstruction(Instr);
1467 Constant *Zero = Builder.getInt32(0);
1468 VectorParts &Entry = WidenMap.get(Instr);
1470 // Handle consecutive loads/stores.
1471 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1472 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1473 setDebugLocFromInst(Builder, Gep);
1474 Value *PtrOperand = Gep->getPointerOperand();
1475 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1476 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1478 // Create the new GEP with the new induction variable.
1479 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1480 Gep2->setOperand(0, FirstBasePtr);
1481 Gep2->setName("gep.indvar.base");
1482 Ptr = Builder.Insert(Gep2);
1484 setDebugLocFromInst(Builder, Gep);
1485 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1486 OrigLoop) && "Base ptr must be invariant");
1488 // The last index does not have to be the induction. It can be
1489 // consecutive and be a function of the index. For example A[I+1];
1490 unsigned NumOperands = Gep->getNumOperands();
1491 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1492 // Create the new GEP with the new induction variable.
1493 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1495 for (unsigned i = 0; i < NumOperands; ++i) {
1496 Value *GepOperand = Gep->getOperand(i);
1497 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1499 // Update last index or loop invariant instruction anchored in loop.
1500 if (i == InductionOperand ||
1501 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1502 assert((i == InductionOperand ||
1503 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1504 "Must be last index or loop invariant");
1506 VectorParts &GEPParts = getVectorValue(GepOperand);
1507 Value *Index = GEPParts[0];
1508 Index = Builder.CreateExtractElement(Index, Zero);
1509 Gep2->setOperand(i, Index);
1510 Gep2->setName("gep.indvar.idx");
1513 Ptr = Builder.Insert(Gep2);
1515 // Use the induction element ptr.
1516 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1517 setDebugLocFromInst(Builder, Ptr);
1518 VectorParts &PtrVal = getVectorValue(Ptr);
1519 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1524 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1525 "We do not allow storing to uniform addresses");
1526 setDebugLocFromInst(Builder, SI);
1527 // We don't want to update the value in the map as it might be used in
1528 // another expression. So don't use a reference type for "StoredVal".
1529 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1531 for (unsigned Part = 0; Part < UF; ++Part) {
1532 // Calculate the pointer for the specific unroll-part.
1533 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1536 // If we store to reverse consecutive memory locations then we need
1537 // to reverse the order of elements in the stored value.
1538 StoredVal[Part] = reverseVector(StoredVal[Part]);
1539 // If the address is consecutive but reversed, then the
1540 // wide store needs to start at the last vector element.
1541 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1542 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1545 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1546 DataTy->getPointerTo(AddressSpace));
1547 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1553 assert(LI && "Must have a load instruction");
1554 setDebugLocFromInst(Builder, LI);
1555 for (unsigned Part = 0; Part < UF; ++Part) {
1556 // Calculate the pointer for the specific unroll-part.
1557 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1560 // If the address is consecutive but reversed, then the
1561 // wide store needs to start at the last vector element.
1562 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1563 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1566 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1567 DataTy->getPointerTo(AddressSpace));
1568 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1569 cast<LoadInst>(LI)->setAlignment(Alignment);
1570 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1574 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1575 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1576 // Holds vector parameters or scalars, in case of uniform vals.
1577 SmallVector<VectorParts, 4> Params;
1579 setDebugLocFromInst(Builder, Instr);
1581 // Find all of the vectorized parameters.
1582 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1583 Value *SrcOp = Instr->getOperand(op);
1585 // If we are accessing the old induction variable, use the new one.
1586 if (SrcOp == OldInduction) {
1587 Params.push_back(getVectorValue(SrcOp));
1591 // Try using previously calculated values.
1592 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1594 // If the src is an instruction that appeared earlier in the basic block
1595 // then it should already be vectorized.
1596 if (SrcInst && OrigLoop->contains(SrcInst)) {
1597 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1598 // The parameter is a vector value from earlier.
1599 Params.push_back(WidenMap.get(SrcInst));
1601 // The parameter is a scalar from outside the loop. Maybe even a constant.
1602 VectorParts Scalars;
1603 Scalars.append(UF, SrcOp);
1604 Params.push_back(Scalars);
1608 assert(Params.size() == Instr->getNumOperands() &&
1609 "Invalid number of operands");
1611 // Does this instruction return a value ?
1612 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1614 Value *UndefVec = IsVoidRetTy ? 0 :
1615 UndefValue::get(VectorType::get(Instr->getType(), VF));
1616 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1617 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1619 Instruction *InsertPt = Builder.GetInsertPoint();
1620 BasicBlock *IfBlock = Builder.GetInsertBlock();
1621 BasicBlock *CondBlock = 0;
1625 if (IfPredicateStore) {
1626 assert(Instr->getParent()->getSinglePredecessor() &&
1627 "Only support single predecessor blocks");
1628 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1629 Instr->getParent());
1630 VectorLp = LI->getLoopFor(IfBlock);
1631 assert(VectorLp && "Must have a loop for this block");
1634 // For each vector unroll 'part':
1635 for (unsigned Part = 0; Part < UF; ++Part) {
1636 // For each scalar that we create:
1637 for (unsigned Width = 0; Width < VF; ++Width) {
1641 if (IfPredicateStore) {
1642 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1643 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1644 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1645 LoopVectorBody.push_back(CondBlock);
1646 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1647 // Update Builder with newly created basic block.
1648 Builder.SetInsertPoint(InsertPt);
1651 Instruction *Cloned = Instr->clone();
1653 Cloned->setName(Instr->getName() + ".cloned");
1654 // Replace the operands of the cloned instructions with extracted scalars.
1655 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1656 Value *Op = Params[op][Part];
1657 // Param is a vector. Need to extract the right lane.
1658 if (Op->getType()->isVectorTy())
1659 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1660 Cloned->setOperand(op, Op);
1663 // Place the cloned scalar in the new loop.
1664 Builder.Insert(Cloned);
1666 // If the original scalar returns a value we need to place it in a vector
1667 // so that future users will be able to use it.
1669 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1670 Builder.getInt32(Width));
1672 if (IfPredicateStore) {
1673 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1674 LoopVectorBody.push_back(NewIfBlock);
1675 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1676 Builder.SetInsertPoint(InsertPt);
1677 Instruction *OldBr = IfBlock->getTerminator();
1678 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1679 OldBr->eraseFromParent();
1680 IfBlock = NewIfBlock;
1686 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1690 if (Instruction *I = dyn_cast<Instruction>(V))
1691 return I->getParent() == Loc->getParent() ? I : 0;
1695 std::pair<Instruction *, Instruction *>
1696 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1697 Instruction *tnullptr = 0;
1698 if (!Legal->mustCheckStrides())
1699 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1701 IRBuilder<> ChkBuilder(Loc);
1705 Instruction *FirstInst = 0;
1706 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1707 SE = Legal->strides_end();
1709 Value *Ptr = stripIntegerCast(*SI);
1710 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1712 // Store the first instruction we create.
1713 FirstInst = getFirstInst(FirstInst, C, Loc);
1715 Check = ChkBuilder.CreateOr(Check, C);
1720 // We have to do this trickery because the IRBuilder might fold the check to a
1721 // constant expression in which case there is no Instruction anchored in a
1723 LLVMContext &Ctx = Loc->getContext();
1724 Instruction *TheCheck =
1725 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1726 ChkBuilder.Insert(TheCheck, "stride.not.one");
1727 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1729 return std::make_pair(FirstInst, TheCheck);
1732 std::pair<Instruction *, Instruction *>
1733 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1734 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1735 Legal->getRuntimePointerCheck();
1737 Instruction *tnullptr = 0;
1738 if (!PtrRtCheck->Need)
1739 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1741 unsigned NumPointers = PtrRtCheck->Pointers.size();
1742 SmallVector<TrackingVH<Value> , 2> Starts;
1743 SmallVector<TrackingVH<Value> , 2> Ends;
1745 LLVMContext &Ctx = Loc->getContext();
1746 SCEVExpander Exp(*SE, "induction");
1747 Instruction *FirstInst = 0;
1749 for (unsigned i = 0; i < NumPointers; ++i) {
1750 Value *Ptr = PtrRtCheck->Pointers[i];
1751 const SCEV *Sc = SE->getSCEV(Ptr);
1753 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1754 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1756 Starts.push_back(Ptr);
1757 Ends.push_back(Ptr);
1759 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1760 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1762 // Use this type for pointer arithmetic.
1763 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1765 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1766 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1767 Starts.push_back(Start);
1768 Ends.push_back(End);
1772 IRBuilder<> ChkBuilder(Loc);
1773 // Our instructions might fold to a constant.
1774 Value *MemoryRuntimeCheck = 0;
1775 for (unsigned i = 0; i < NumPointers; ++i) {
1776 for (unsigned j = i+1; j < NumPointers; ++j) {
1777 // No need to check if two readonly pointers intersect.
1778 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1781 // Only need to check pointers between two different dependency sets.
1782 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1785 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1786 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1788 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1789 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1790 "Trying to bounds check pointers with different address spaces");
1792 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1793 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1795 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1796 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1797 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1798 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1800 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1801 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1802 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1803 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1804 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1805 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1806 if (MemoryRuntimeCheck) {
1807 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1809 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1811 MemoryRuntimeCheck = IsConflict;
1815 // We have to do this trickery because the IRBuilder might fold the check to a
1816 // constant expression in which case there is no Instruction anchored in a
1818 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1819 ConstantInt::getTrue(Ctx));
1820 ChkBuilder.Insert(Check, "memcheck.conflict");
1821 FirstInst = getFirstInst(FirstInst, Check, Loc);
1822 return std::make_pair(FirstInst, Check);
1825 void InnerLoopVectorizer::createEmptyLoop() {
1827 In this function we generate a new loop. The new loop will contain
1828 the vectorized instructions while the old loop will continue to run the
1831 [ ] <-- vector loop bypass (may consist of multiple blocks).
1834 | [ ] <-- vector pre header.
1838 | [ ]_| <-- vector loop.
1841 >[ ] <--- middle-block.
1844 | [ ] <--- new preheader.
1848 | [ ]_| <-- old scalar loop to handle remainder.
1851 >[ ] <-- exit block.
1855 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1856 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1857 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1858 assert(ExitBlock && "Must have an exit block");
1860 // Some loops have a single integer induction variable, while other loops
1861 // don't. One example is c++ iterators that often have multiple pointer
1862 // induction variables. In the code below we also support a case where we
1863 // don't have a single induction variable.
1864 OldInduction = Legal->getInduction();
1865 Type *IdxTy = Legal->getWidestInductionType();
1867 // Find the loop boundaries.
1868 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1869 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1871 // The exit count might have the type of i64 while the phi is i32. This can
1872 // happen if we have an induction variable that is sign extended before the
1873 // compare. The only way that we get a backedge taken count is that the
1874 // induction variable was signed and as such will not overflow. In such a case
1875 // truncation is legal.
1876 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1877 IdxTy->getPrimitiveSizeInBits())
1878 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1880 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1881 // Get the total trip count from the count by adding 1.
1882 ExitCount = SE->getAddExpr(ExitCount,
1883 SE->getConstant(ExitCount->getType(), 1));
1885 // Expand the trip count and place the new instructions in the preheader.
1886 // Notice that the pre-header does not change, only the loop body.
1887 SCEVExpander Exp(*SE, "induction");
1889 // Count holds the overall loop count (N).
1890 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1891 BypassBlock->getTerminator());
1893 // The loop index does not have to start at Zero. Find the original start
1894 // value from the induction PHI node. If we don't have an induction variable
1895 // then we know that it starts at zero.
1896 Builder.SetInsertPoint(BypassBlock->getTerminator());
1897 Value *StartIdx = ExtendedIdx = OldInduction ?
1898 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1900 ConstantInt::get(IdxTy, 0);
1902 assert(BypassBlock && "Invalid loop structure");
1903 LoopBypassBlocks.push_back(BypassBlock);
1905 // Split the single block loop into the two loop structure described above.
1906 BasicBlock *VectorPH =
1907 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1908 BasicBlock *VecBody =
1909 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1910 BasicBlock *MiddleBlock =
1911 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1912 BasicBlock *ScalarPH =
1913 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1915 // Create and register the new vector loop.
1916 Loop* Lp = new Loop();
1917 Loop *ParentLoop = OrigLoop->getParentLoop();
1919 // Insert the new loop into the loop nest and register the new basic blocks
1920 // before calling any utilities such as SCEV that require valid LoopInfo.
1922 ParentLoop->addChildLoop(Lp);
1923 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1924 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1925 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1927 LI->addTopLevelLoop(Lp);
1929 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1931 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1933 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1935 // Generate the induction variable.
1936 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1937 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1938 // The loop step is equal to the vectorization factor (num of SIMD elements)
1939 // times the unroll factor (num of SIMD instructions).
1940 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1942 // This is the IR builder that we use to add all of the logic for bypassing
1943 // the new vector loop.
1944 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1945 setDebugLocFromInst(BypassBuilder,
1946 getDebugLocFromInstOrOperands(OldInduction));
1948 // We may need to extend the index in case there is a type mismatch.
1949 // We know that the count starts at zero and does not overflow.
1950 if (Count->getType() != IdxTy) {
1951 // The exit count can be of pointer type. Convert it to the correct
1953 if (ExitCount->getType()->isPointerTy())
1954 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1956 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1959 // Add the start index to the loop count to get the new end index.
1960 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1962 // Now we need to generate the expression for N - (N % VF), which is
1963 // the part that the vectorized body will execute.
1964 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1965 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1966 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1967 "end.idx.rnd.down");
1969 // Now, compare the new count to zero. If it is zero skip the vector loop and
1970 // jump to the scalar loop.
1971 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1974 BasicBlock *LastBypassBlock = BypassBlock;
1976 // Generate the code to check that the strides we assumed to be one are really
1977 // one. We want the new basic block to start at the first instruction in a
1978 // sequence of instructions that form a check.
1979 Instruction *StrideCheck;
1980 Instruction *FirstCheckInst;
1981 std::tie(FirstCheckInst, StrideCheck) =
1982 addStrideCheck(BypassBlock->getTerminator());
1984 // Create a new block containing the stride check.
1985 BasicBlock *CheckBlock =
1986 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
1988 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1989 LoopBypassBlocks.push_back(CheckBlock);
1991 // Replace the branch into the memory check block with a conditional branch
1992 // for the "few elements case".
1993 Instruction *OldTerm = BypassBlock->getTerminator();
1994 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1995 OldTerm->eraseFromParent();
1998 LastBypassBlock = CheckBlock;
2001 // Generate the code that checks in runtime if arrays overlap. We put the
2002 // checks into a separate block to make the more common case of few elements
2004 Instruction *MemRuntimeCheck;
2005 std::tie(FirstCheckInst, MemRuntimeCheck) =
2006 addRuntimeCheck(LastBypassBlock->getTerminator());
2007 if (MemRuntimeCheck) {
2008 // Create a new block containing the memory check.
2009 BasicBlock *CheckBlock =
2010 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2012 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2013 LoopBypassBlocks.push_back(CheckBlock);
2015 // Replace the branch into the memory check block with a conditional branch
2016 // for the "few elements case".
2017 Instruction *OldTerm = LastBypassBlock->getTerminator();
2018 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2019 OldTerm->eraseFromParent();
2021 Cmp = MemRuntimeCheck;
2022 LastBypassBlock = CheckBlock;
2025 LastBypassBlock->getTerminator()->eraseFromParent();
2026 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2029 // We are going to resume the execution of the scalar loop.
2030 // Go over all of the induction variables that we found and fix the
2031 // PHIs that are left in the scalar version of the loop.
2032 // The starting values of PHI nodes depend on the counter of the last
2033 // iteration in the vectorized loop.
2034 // If we come from a bypass edge then we need to start from the original
2037 // This variable saves the new starting index for the scalar loop.
2038 PHINode *ResumeIndex = 0;
2039 LoopVectorizationLegality::InductionList::iterator I, E;
2040 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2041 // Set builder to point to last bypass block.
2042 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2043 for (I = List->begin(), E = List->end(); I != E; ++I) {
2044 PHINode *OrigPhi = I->first;
2045 LoopVectorizationLegality::InductionInfo II = I->second;
2047 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2048 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2049 MiddleBlock->getTerminator());
2050 // We might have extended the type of the induction variable but we need a
2051 // truncated version for the scalar loop.
2052 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2053 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2054 MiddleBlock->getTerminator()) : 0;
2056 Value *EndValue = 0;
2058 case LoopVectorizationLegality::IK_NoInduction:
2059 llvm_unreachable("Unknown induction");
2060 case LoopVectorizationLegality::IK_IntInduction: {
2061 // Handle the integer induction counter.
2062 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2064 // We have the canonical induction variable.
2065 if (OrigPhi == OldInduction) {
2066 // Create a truncated version of the resume value for the scalar loop,
2067 // we might have promoted the type to a larger width.
2069 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2070 // The new PHI merges the original incoming value, in case of a bypass,
2071 // or the value at the end of the vectorized loop.
2072 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2073 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2074 TruncResumeVal->addIncoming(EndValue, VecBody);
2076 // We know what the end value is.
2077 EndValue = IdxEndRoundDown;
2078 // We also know which PHI node holds it.
2079 ResumeIndex = ResumeVal;
2083 // Not the canonical induction variable - add the vector loop count to the
2085 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2086 II.StartValue->getType(),
2088 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2091 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2092 // Convert the CountRoundDown variable to the PHI size.
2093 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2094 II.StartValue->getType(),
2096 // Handle reverse integer induction counter.
2097 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2100 case LoopVectorizationLegality::IK_PtrInduction: {
2101 // For pointer induction variables, calculate the offset using
2103 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2107 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2108 // The value at the end of the loop for the reverse pointer is calculated
2109 // by creating a GEP with a negative index starting from the start value.
2110 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2111 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2113 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2119 // The new PHI merges the original incoming value, in case of a bypass,
2120 // or the value at the end of the vectorized loop.
2121 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2122 if (OrigPhi == OldInduction)
2123 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2125 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2127 ResumeVal->addIncoming(EndValue, VecBody);
2129 // Fix the scalar body counter (PHI node).
2130 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2131 // The old inductions phi node in the scalar body needs the truncated value.
2132 if (OrigPhi == OldInduction)
2133 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2135 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2138 // If we are generating a new induction variable then we also need to
2139 // generate the code that calculates the exit value. This value is not
2140 // simply the end of the counter because we may skip the vectorized body
2141 // in case of a runtime check.
2143 assert(!ResumeIndex && "Unexpected resume value found");
2144 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2145 MiddleBlock->getTerminator());
2146 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2147 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2148 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2151 // Make sure that we found the index where scalar loop needs to continue.
2152 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2153 "Invalid resume Index");
2155 // Add a check in the middle block to see if we have completed
2156 // all of the iterations in the first vector loop.
2157 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2158 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2159 ResumeIndex, "cmp.n",
2160 MiddleBlock->getTerminator());
2162 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2163 // Remove the old terminator.
2164 MiddleBlock->getTerminator()->eraseFromParent();
2166 // Create i+1 and fill the PHINode.
2167 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2168 Induction->addIncoming(StartIdx, VectorPH);
2169 Induction->addIncoming(NextIdx, VecBody);
2170 // Create the compare.
2171 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2172 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2174 // Now we have two terminators. Remove the old one from the block.
2175 VecBody->getTerminator()->eraseFromParent();
2177 // Get ready to start creating new instructions into the vectorized body.
2178 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2181 LoopVectorPreHeader = VectorPH;
2182 LoopScalarPreHeader = ScalarPH;
2183 LoopMiddleBlock = MiddleBlock;
2184 LoopExitBlock = ExitBlock;
2185 LoopVectorBody.push_back(VecBody);
2186 LoopScalarBody = OldBasicBlock;
2188 LoopVectorizeHints Hints(Lp, true);
2189 Hints.setAlreadyVectorized(Lp);
2192 /// This function returns the identity element (or neutral element) for
2193 /// the operation K.
2195 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2200 // Adding, Xoring, Oring zero to a number does not change it.
2201 return ConstantInt::get(Tp, 0);
2202 case RK_IntegerMult:
2203 // Multiplying a number by 1 does not change it.
2204 return ConstantInt::get(Tp, 1);
2206 // AND-ing a number with an all-1 value does not change it.
2207 return ConstantInt::get(Tp, -1, true);
2209 // Multiplying a number by 1 does not change it.
2210 return ConstantFP::get(Tp, 1.0L);
2212 // Adding zero to a number does not change it.
2213 return ConstantFP::get(Tp, 0.0L);
2215 llvm_unreachable("Unknown reduction kind");
2219 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2220 Intrinsic::ID ValidIntrinsicID) {
2221 if (I.getNumArgOperands() != 1 ||
2222 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2223 I.getType() != I.getArgOperand(0)->getType() ||
2224 !I.onlyReadsMemory())
2225 return Intrinsic::not_intrinsic;
2227 return ValidIntrinsicID;
2230 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2231 Intrinsic::ID ValidIntrinsicID) {
2232 if (I.getNumArgOperands() != 2 ||
2233 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2234 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2235 I.getType() != I.getArgOperand(0)->getType() ||
2236 I.getType() != I.getArgOperand(1)->getType() ||
2237 !I.onlyReadsMemory())
2238 return Intrinsic::not_intrinsic;
2240 return ValidIntrinsicID;
2244 static Intrinsic::ID
2245 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2246 // If we have an intrinsic call, check if it is trivially vectorizable.
2247 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2248 switch (II->getIntrinsicID()) {
2249 case Intrinsic::sqrt:
2250 case Intrinsic::sin:
2251 case Intrinsic::cos:
2252 case Intrinsic::exp:
2253 case Intrinsic::exp2:
2254 case Intrinsic::log:
2255 case Intrinsic::log10:
2256 case Intrinsic::log2:
2257 case Intrinsic::fabs:
2258 case Intrinsic::copysign:
2259 case Intrinsic::floor:
2260 case Intrinsic::ceil:
2261 case Intrinsic::trunc:
2262 case Intrinsic::rint:
2263 case Intrinsic::nearbyint:
2264 case Intrinsic::round:
2265 case Intrinsic::pow:
2266 case Intrinsic::fma:
2267 case Intrinsic::fmuladd:
2268 case Intrinsic::lifetime_start:
2269 case Intrinsic::lifetime_end:
2270 return II->getIntrinsicID();
2272 return Intrinsic::not_intrinsic;
2277 return Intrinsic::not_intrinsic;
2280 Function *F = CI->getCalledFunction();
2281 // We're going to make assumptions on the semantics of the functions, check
2282 // that the target knows that it's available in this environment and it does
2283 // not have local linkage.
2284 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2285 return Intrinsic::not_intrinsic;
2287 // Otherwise check if we have a call to a function that can be turned into a
2288 // vector intrinsic.
2295 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2299 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2303 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2305 case LibFunc::exp2f:
2306 case LibFunc::exp2l:
2307 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2311 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2312 case LibFunc::log10:
2313 case LibFunc::log10f:
2314 case LibFunc::log10l:
2315 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2317 case LibFunc::log2f:
2318 case LibFunc::log2l:
2319 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2321 case LibFunc::fabsf:
2322 case LibFunc::fabsl:
2323 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2324 case LibFunc::copysign:
2325 case LibFunc::copysignf:
2326 case LibFunc::copysignl:
2327 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2328 case LibFunc::floor:
2329 case LibFunc::floorf:
2330 case LibFunc::floorl:
2331 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2333 case LibFunc::ceilf:
2334 case LibFunc::ceill:
2335 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2336 case LibFunc::trunc:
2337 case LibFunc::truncf:
2338 case LibFunc::truncl:
2339 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2341 case LibFunc::rintf:
2342 case LibFunc::rintl:
2343 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2344 case LibFunc::nearbyint:
2345 case LibFunc::nearbyintf:
2346 case LibFunc::nearbyintl:
2347 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2348 case LibFunc::round:
2349 case LibFunc::roundf:
2350 case LibFunc::roundl:
2351 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2355 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2358 return Intrinsic::not_intrinsic;
2361 /// This function translates the reduction kind to an LLVM binary operator.
2363 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2365 case LoopVectorizationLegality::RK_IntegerAdd:
2366 return Instruction::Add;
2367 case LoopVectorizationLegality::RK_IntegerMult:
2368 return Instruction::Mul;
2369 case LoopVectorizationLegality::RK_IntegerOr:
2370 return Instruction::Or;
2371 case LoopVectorizationLegality::RK_IntegerAnd:
2372 return Instruction::And;
2373 case LoopVectorizationLegality::RK_IntegerXor:
2374 return Instruction::Xor;
2375 case LoopVectorizationLegality::RK_FloatMult:
2376 return Instruction::FMul;
2377 case LoopVectorizationLegality::RK_FloatAdd:
2378 return Instruction::FAdd;
2379 case LoopVectorizationLegality::RK_IntegerMinMax:
2380 return Instruction::ICmp;
2381 case LoopVectorizationLegality::RK_FloatMinMax:
2382 return Instruction::FCmp;
2384 llvm_unreachable("Unknown reduction operation");
2388 Value *createMinMaxOp(IRBuilder<> &Builder,
2389 LoopVectorizationLegality::MinMaxReductionKind RK,
2392 CmpInst::Predicate P = CmpInst::ICMP_NE;
2395 llvm_unreachable("Unknown min/max reduction kind");
2396 case LoopVectorizationLegality::MRK_UIntMin:
2397 P = CmpInst::ICMP_ULT;
2399 case LoopVectorizationLegality::MRK_UIntMax:
2400 P = CmpInst::ICMP_UGT;
2402 case LoopVectorizationLegality::MRK_SIntMin:
2403 P = CmpInst::ICMP_SLT;
2405 case LoopVectorizationLegality::MRK_SIntMax:
2406 P = CmpInst::ICMP_SGT;
2408 case LoopVectorizationLegality::MRK_FloatMin:
2409 P = CmpInst::FCMP_OLT;
2411 case LoopVectorizationLegality::MRK_FloatMax:
2412 P = CmpInst::FCMP_OGT;
2417 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2418 RK == LoopVectorizationLegality::MRK_FloatMax)
2419 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2421 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2423 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2428 struct CSEDenseMapInfo {
2429 static bool canHandle(Instruction *I) {
2430 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2431 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2433 static inline Instruction *getEmptyKey() {
2434 return DenseMapInfo<Instruction *>::getEmptyKey();
2436 static inline Instruction *getTombstoneKey() {
2437 return DenseMapInfo<Instruction *>::getTombstoneKey();
2439 static unsigned getHashValue(Instruction *I) {
2440 assert(canHandle(I) && "Unknown instruction!");
2441 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2442 I->value_op_end()));
2444 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2445 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2446 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2448 return LHS->isIdenticalTo(RHS);
2453 /// \brief Check whether this block is a predicated block.
2454 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2455 /// = ...; " blocks. We start with one vectorized basic block. For every
2456 /// conditional block we split this vectorized block. Therefore, every second
2457 /// block will be a predicated one.
2458 static bool isPredicatedBlock(unsigned BlockNum) {
2459 return BlockNum % 2;
2462 ///\brief Perform cse of induction variable instructions.
2463 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2464 // Perform simple cse.
2465 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2466 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2467 BasicBlock *BB = BBs[i];
2468 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2469 Instruction *In = I++;
2471 if (!CSEDenseMapInfo::canHandle(In))
2474 // Check if we can replace this instruction with any of the
2475 // visited instructions.
2476 if (Instruction *V = CSEMap.lookup(In)) {
2477 In->replaceAllUsesWith(V);
2478 In->eraseFromParent();
2481 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2482 // ...;" blocks for predicated stores. Every second block is a predicated
2484 if (isPredicatedBlock(i))
2492 void InnerLoopVectorizer::vectorizeLoop() {
2493 //===------------------------------------------------===//
2495 // Notice: any optimization or new instruction that go
2496 // into the code below should be also be implemented in
2499 //===------------------------------------------------===//
2500 Constant *Zero = Builder.getInt32(0);
2502 // In order to support reduction variables we need to be able to vectorize
2503 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2504 // stages. First, we create a new vector PHI node with no incoming edges.
2505 // We use this value when we vectorize all of the instructions that use the
2506 // PHI. Next, after all of the instructions in the block are complete we
2507 // add the new incoming edges to the PHI. At this point all of the
2508 // instructions in the basic block are vectorized, so we can use them to
2509 // construct the PHI.
2510 PhiVector RdxPHIsToFix;
2512 // Scan the loop in a topological order to ensure that defs are vectorized
2514 LoopBlocksDFS DFS(OrigLoop);
2517 // Vectorize all of the blocks in the original loop.
2518 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2519 be = DFS.endRPO(); bb != be; ++bb)
2520 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2522 // At this point every instruction in the original loop is widened to
2523 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2524 // that we vectorized. The PHI nodes are currently empty because we did
2525 // not want to introduce cycles. Notice that the remaining PHI nodes
2526 // that we need to fix are reduction variables.
2528 // Create the 'reduced' values for each of the induction vars.
2529 // The reduced values are the vector values that we scalarize and combine
2530 // after the loop is finished.
2531 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2533 PHINode *RdxPhi = *it;
2534 assert(RdxPhi && "Unable to recover vectorized PHI");
2536 // Find the reduction variable descriptor.
2537 assert(Legal->getReductionVars()->count(RdxPhi) &&
2538 "Unable to find the reduction variable");
2539 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2540 (*Legal->getReductionVars())[RdxPhi];
2542 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2544 // We need to generate a reduction vector from the incoming scalar.
2545 // To do so, we need to generate the 'identity' vector and override
2546 // one of the elements with the incoming scalar reduction. We need
2547 // to do it in the vector-loop preheader.
2548 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2550 // This is the vector-clone of the value that leaves the loop.
2551 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2552 Type *VecTy = VectorExit[0]->getType();
2554 // Find the reduction identity variable. Zero for addition, or, xor,
2555 // one for multiplication, -1 for And.
2558 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2559 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2560 // MinMax reduction have the start value as their identify.
2562 VectorStart = Identity = RdxDesc.StartValue;
2564 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2569 // Handle other reduction kinds:
2571 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2572 VecTy->getScalarType());
2575 // This vector is the Identity vector where the first element is the
2576 // incoming scalar reduction.
2577 VectorStart = RdxDesc.StartValue;
2579 Identity = ConstantVector::getSplat(VF, Iden);
2581 // This vector is the Identity vector where the first element is the
2582 // incoming scalar reduction.
2583 VectorStart = Builder.CreateInsertElement(Identity,
2584 RdxDesc.StartValue, Zero);
2588 // Fix the vector-loop phi.
2589 // We created the induction variable so we know that the
2590 // preheader is the first entry.
2591 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2593 // Reductions do not have to start at zero. They can start with
2594 // any loop invariant values.
2595 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2596 BasicBlock *Latch = OrigLoop->getLoopLatch();
2597 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2598 VectorParts &Val = getVectorValue(LoopVal);
2599 for (unsigned part = 0; part < UF; ++part) {
2600 // Make sure to add the reduction stat value only to the
2601 // first unroll part.
2602 Value *StartVal = (part == 0) ? VectorStart : Identity;
2603 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2604 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2605 LoopVectorBody.back());
2608 // Before each round, move the insertion point right between
2609 // the PHIs and the values we are going to write.
2610 // This allows us to write both PHINodes and the extractelement
2612 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2614 VectorParts RdxParts;
2615 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2616 for (unsigned part = 0; part < UF; ++part) {
2617 // This PHINode contains the vectorized reduction variable, or
2618 // the initial value vector, if we bypass the vector loop.
2619 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2620 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2621 Value *StartVal = (part == 0) ? VectorStart : Identity;
2622 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2623 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2624 NewPhi->addIncoming(RdxExitVal[part],
2625 LoopVectorBody.back());
2626 RdxParts.push_back(NewPhi);
2629 // Reduce all of the unrolled parts into a single vector.
2630 Value *ReducedPartRdx = RdxParts[0];
2631 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2632 setDebugLocFromInst(Builder, ReducedPartRdx);
2633 for (unsigned part = 1; part < UF; ++part) {
2634 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2635 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2636 RdxParts[part], ReducedPartRdx,
2639 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2640 ReducedPartRdx, RdxParts[part]);
2644 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2645 // and vector ops, reducing the set of values being computed by half each
2647 assert(isPowerOf2_32(VF) &&
2648 "Reduction emission only supported for pow2 vectors!");
2649 Value *TmpVec = ReducedPartRdx;
2650 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2651 for (unsigned i = VF; i != 1; i >>= 1) {
2652 // Move the upper half of the vector to the lower half.
2653 for (unsigned j = 0; j != i/2; ++j)
2654 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2656 // Fill the rest of the mask with undef.
2657 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2658 UndefValue::get(Builder.getInt32Ty()));
2661 Builder.CreateShuffleVector(TmpVec,
2662 UndefValue::get(TmpVec->getType()),
2663 ConstantVector::get(ShuffleMask),
2666 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2667 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2670 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2673 // The result is in the first element of the vector.
2674 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2675 Builder.getInt32(0));
2678 // Now, we need to fix the users of the reduction variable
2679 // inside and outside of the scalar remainder loop.
2680 // We know that the loop is in LCSSA form. We need to update the
2681 // PHI nodes in the exit blocks.
2682 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2683 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2684 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2685 if (!LCSSAPhi) break;
2687 // All PHINodes need to have a single entry edge, or two if
2688 // we already fixed them.
2689 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2691 // We found our reduction value exit-PHI. Update it with the
2692 // incoming bypass edge.
2693 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2694 // Add an edge coming from the bypass.
2695 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2698 }// end of the LCSSA phi scan.
2700 // Fix the scalar loop reduction variable with the incoming reduction sum
2701 // from the vector body and from the backedge value.
2702 int IncomingEdgeBlockIdx =
2703 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2704 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2705 // Pick the other block.
2706 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2707 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2708 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2709 }// end of for each redux variable.
2713 // Remove redundant induction instructions.
2714 cse(LoopVectorBody);
2717 void InnerLoopVectorizer::fixLCSSAPHIs() {
2718 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2719 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2720 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2721 if (!LCSSAPhi) break;
2722 if (LCSSAPhi->getNumIncomingValues() == 1)
2723 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2728 InnerLoopVectorizer::VectorParts
2729 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2730 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2733 // Look for cached value.
2734 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2735 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2736 if (ECEntryIt != MaskCache.end())
2737 return ECEntryIt->second;
2739 VectorParts SrcMask = createBlockInMask(Src);
2741 // The terminator has to be a branch inst!
2742 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2743 assert(BI && "Unexpected terminator found");
2745 if (BI->isConditional()) {
2746 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2748 if (BI->getSuccessor(0) != Dst)
2749 for (unsigned part = 0; part < UF; ++part)
2750 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2752 for (unsigned part = 0; part < UF; ++part)
2753 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2755 MaskCache[Edge] = EdgeMask;
2759 MaskCache[Edge] = SrcMask;
2763 InnerLoopVectorizer::VectorParts
2764 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2765 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2767 // Loop incoming mask is all-one.
2768 if (OrigLoop->getHeader() == BB) {
2769 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2770 return getVectorValue(C);
2773 // This is the block mask. We OR all incoming edges, and with zero.
2774 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2775 VectorParts BlockMask = getVectorValue(Zero);
2778 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2779 VectorParts EM = createEdgeMask(*it, BB);
2780 for (unsigned part = 0; part < UF; ++part)
2781 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2787 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2788 InnerLoopVectorizer::VectorParts &Entry,
2789 unsigned UF, unsigned VF, PhiVector *PV) {
2790 PHINode* P = cast<PHINode>(PN);
2791 // Handle reduction variables:
2792 if (Legal->getReductionVars()->count(P)) {
2793 for (unsigned part = 0; part < UF; ++part) {
2794 // This is phase one of vectorizing PHIs.
2795 Type *VecTy = (VF == 1) ? PN->getType() :
2796 VectorType::get(PN->getType(), VF);
2797 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2798 LoopVectorBody.back()-> getFirstInsertionPt());
2804 setDebugLocFromInst(Builder, P);
2805 // Check for PHI nodes that are lowered to vector selects.
2806 if (P->getParent() != OrigLoop->getHeader()) {
2807 // We know that all PHIs in non-header blocks are converted into
2808 // selects, so we don't have to worry about the insertion order and we
2809 // can just use the builder.
2810 // At this point we generate the predication tree. There may be
2811 // duplications since this is a simple recursive scan, but future
2812 // optimizations will clean it up.
2814 unsigned NumIncoming = P->getNumIncomingValues();
2816 // Generate a sequence of selects of the form:
2817 // SELECT(Mask3, In3,
2818 // SELECT(Mask2, In2,
2820 for (unsigned In = 0; In < NumIncoming; In++) {
2821 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2823 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2825 for (unsigned part = 0; part < UF; ++part) {
2826 // We might have single edge PHIs (blocks) - use an identity
2827 // 'select' for the first PHI operand.
2829 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2832 // Select between the current value and the previous incoming edge
2833 // based on the incoming mask.
2834 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2835 Entry[part], "predphi");
2841 // This PHINode must be an induction variable.
2842 // Make sure that we know about it.
2843 assert(Legal->getInductionVars()->count(P) &&
2844 "Not an induction variable");
2846 LoopVectorizationLegality::InductionInfo II =
2847 Legal->getInductionVars()->lookup(P);
2850 case LoopVectorizationLegality::IK_NoInduction:
2851 llvm_unreachable("Unknown induction");
2852 case LoopVectorizationLegality::IK_IntInduction: {
2853 assert(P->getType() == II.StartValue->getType() && "Types must match");
2854 Type *PhiTy = P->getType();
2856 if (P == OldInduction) {
2857 // Handle the canonical induction variable. We might have had to
2859 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2861 // Handle other induction variables that are now based on the
2863 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2865 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2866 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2869 Broadcasted = getBroadcastInstrs(Broadcasted);
2870 // After broadcasting the induction variable we need to make the vector
2871 // consecutive by adding 0, 1, 2, etc.
2872 for (unsigned part = 0; part < UF; ++part)
2873 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2876 case LoopVectorizationLegality::IK_ReverseIntInduction:
2877 case LoopVectorizationLegality::IK_PtrInduction:
2878 case LoopVectorizationLegality::IK_ReversePtrInduction:
2879 // Handle reverse integer and pointer inductions.
2880 Value *StartIdx = ExtendedIdx;
2881 // This is the normalized GEP that starts counting at zero.
2882 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2885 // Handle the reverse integer induction variable case.
2886 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2887 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2888 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2890 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2893 // This is a new value so do not hoist it out.
2894 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2895 // After broadcasting the induction variable we need to make the
2896 // vector consecutive by adding ... -3, -2, -1, 0.
2897 for (unsigned part = 0; part < UF; ++part)
2898 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2903 // Handle the pointer induction variable case.
2904 assert(P->getType()->isPointerTy() && "Unexpected type.");
2906 // Is this a reverse induction ptr or a consecutive induction ptr.
2907 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2910 // This is the vector of results. Notice that we don't generate
2911 // vector geps because scalar geps result in better code.
2912 for (unsigned part = 0; part < UF; ++part) {
2914 int EltIndex = (part) * (Reverse ? -1 : 1);
2915 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2918 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2920 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2922 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2924 Entry[part] = SclrGep;
2928 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2929 for (unsigned int i = 0; i < VF; ++i) {
2930 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2931 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2934 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2936 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2938 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2940 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2941 Builder.getInt32(i),
2944 Entry[part] = VecVal;
2950 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2951 // For each instruction in the old loop.
2952 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2953 VectorParts &Entry = WidenMap.get(it);
2954 switch (it->getOpcode()) {
2955 case Instruction::Br:
2956 // Nothing to do for PHIs and BR, since we already took care of the
2957 // loop control flow instructions.
2959 case Instruction::PHI:{
2960 // Vectorize PHINodes.
2961 widenPHIInstruction(it, Entry, UF, VF, PV);
2965 case Instruction::Add:
2966 case Instruction::FAdd:
2967 case Instruction::Sub:
2968 case Instruction::FSub:
2969 case Instruction::Mul:
2970 case Instruction::FMul:
2971 case Instruction::UDiv:
2972 case Instruction::SDiv:
2973 case Instruction::FDiv:
2974 case Instruction::URem:
2975 case Instruction::SRem:
2976 case Instruction::FRem:
2977 case Instruction::Shl:
2978 case Instruction::LShr:
2979 case Instruction::AShr:
2980 case Instruction::And:
2981 case Instruction::Or:
2982 case Instruction::Xor: {
2983 // Just widen binops.
2984 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2985 setDebugLocFromInst(Builder, BinOp);
2986 VectorParts &A = getVectorValue(it->getOperand(0));
2987 VectorParts &B = getVectorValue(it->getOperand(1));
2989 // Use this vector value for all users of the original instruction.
2990 for (unsigned Part = 0; Part < UF; ++Part) {
2991 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2993 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2994 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2995 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2996 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2997 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2999 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3000 VecOp->setIsExact(BinOp->isExact());
3006 case Instruction::Select: {
3008 // If the selector is loop invariant we can create a select
3009 // instruction with a scalar condition. Otherwise, use vector-select.
3010 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3012 setDebugLocFromInst(Builder, it);
3014 // The condition can be loop invariant but still defined inside the
3015 // loop. This means that we can't just use the original 'cond' value.
3016 // We have to take the 'vectorized' value and pick the first lane.
3017 // Instcombine will make this a no-op.
3018 VectorParts &Cond = getVectorValue(it->getOperand(0));
3019 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3020 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3022 Value *ScalarCond = (VF == 1) ? Cond[0] :
3023 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3025 for (unsigned Part = 0; Part < UF; ++Part) {
3026 Entry[Part] = Builder.CreateSelect(
3027 InvariantCond ? ScalarCond : Cond[Part],
3034 case Instruction::ICmp:
3035 case Instruction::FCmp: {
3036 // Widen compares. Generate vector compares.
3037 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3038 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3039 setDebugLocFromInst(Builder, it);
3040 VectorParts &A = getVectorValue(it->getOperand(0));
3041 VectorParts &B = getVectorValue(it->getOperand(1));
3042 for (unsigned Part = 0; Part < UF; ++Part) {
3045 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3047 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3053 case Instruction::Store:
3054 case Instruction::Load:
3055 vectorizeMemoryInstruction(it);
3057 case Instruction::ZExt:
3058 case Instruction::SExt:
3059 case Instruction::FPToUI:
3060 case Instruction::FPToSI:
3061 case Instruction::FPExt:
3062 case Instruction::PtrToInt:
3063 case Instruction::IntToPtr:
3064 case Instruction::SIToFP:
3065 case Instruction::UIToFP:
3066 case Instruction::Trunc:
3067 case Instruction::FPTrunc:
3068 case Instruction::BitCast: {
3069 CastInst *CI = dyn_cast<CastInst>(it);
3070 setDebugLocFromInst(Builder, it);
3071 /// Optimize the special case where the source is the induction
3072 /// variable. Notice that we can only optimize the 'trunc' case
3073 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3074 /// c. other casts depend on pointer size.
3075 if (CI->getOperand(0) == OldInduction &&
3076 it->getOpcode() == Instruction::Trunc) {
3077 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3079 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3080 for (unsigned Part = 0; Part < UF; ++Part)
3081 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3084 /// Vectorize casts.
3085 Type *DestTy = (VF == 1) ? CI->getType() :
3086 VectorType::get(CI->getType(), VF);
3088 VectorParts &A = getVectorValue(it->getOperand(0));
3089 for (unsigned Part = 0; Part < UF; ++Part)
3090 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3094 case Instruction::Call: {
3095 // Ignore dbg intrinsics.
3096 if (isa<DbgInfoIntrinsic>(it))
3098 setDebugLocFromInst(Builder, it);
3100 Module *M = BB->getParent()->getParent();
3101 CallInst *CI = cast<CallInst>(it);
3102 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3103 assert(ID && "Not an intrinsic call!");
3105 case Intrinsic::lifetime_end:
3106 case Intrinsic::lifetime_start:
3107 scalarizeInstruction(it);
3110 for (unsigned Part = 0; Part < UF; ++Part) {
3111 SmallVector<Value *, 4> Args;
3112 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3113 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3114 Args.push_back(Arg[Part]);
3116 Type *Tys[] = {CI->getType()};
3118 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3120 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3121 Entry[Part] = Builder.CreateCall(F, Args);
3129 // All other instructions are unsupported. Scalarize them.
3130 scalarizeInstruction(it);
3133 }// end of for_each instr.
3136 void InnerLoopVectorizer::updateAnalysis() {
3137 // Forget the original basic block.
3138 SE->forgetLoop(OrigLoop);
3140 // Update the dominator tree information.
3141 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3142 "Entry does not dominate exit.");
3144 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3145 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3146 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3148 // Due to if predication of stores we might create a sequence of "if(pred)
3149 // a[i] = ...; " blocks.
3150 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3152 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3153 else if (isPredicatedBlock(i)) {
3154 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3156 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3160 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3161 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3162 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3163 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3165 DEBUG(DT->verifyDomTree());
3168 /// \brief Check whether it is safe to if-convert this phi node.
3170 /// Phi nodes with constant expressions that can trap are not safe to if
3172 static bool canIfConvertPHINodes(BasicBlock *BB) {
3173 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3174 PHINode *Phi = dyn_cast<PHINode>(I);
3177 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3178 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3185 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3186 if (!EnableIfConversion)
3189 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3191 // A list of pointers that we can safely read and write to.
3192 SmallPtrSet<Value *, 8> SafePointes;
3194 // Collect safe addresses.
3195 for (Loop::block_iterator BI = TheLoop->block_begin(),
3196 BE = TheLoop->block_end(); BI != BE; ++BI) {
3197 BasicBlock *BB = *BI;
3199 if (blockNeedsPredication(BB))
3202 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3203 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3204 SafePointes.insert(LI->getPointerOperand());
3205 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3206 SafePointes.insert(SI->getPointerOperand());
3210 // Collect the blocks that need predication.
3211 BasicBlock *Header = TheLoop->getHeader();
3212 for (Loop::block_iterator BI = TheLoop->block_begin(),
3213 BE = TheLoop->block_end(); BI != BE; ++BI) {
3214 BasicBlock *BB = *BI;
3216 // We don't support switch statements inside loops.
3217 if (!isa<BranchInst>(BB->getTerminator()))
3220 // We must be able to predicate all blocks that need to be predicated.
3221 if (blockNeedsPredication(BB)) {
3222 if (!blockCanBePredicated(BB, SafePointes))
3224 } else if (BB != Header && !canIfConvertPHINodes(BB))
3229 // We can if-convert this loop.
3233 bool LoopVectorizationLegality::canVectorize() {
3234 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3235 // be canonicalized.
3236 if (!TheLoop->getLoopPreheader())
3239 // We can only vectorize innermost loops.
3240 if (TheLoop->getSubLoopsVector().size())
3243 // We must have a single backedge.
3244 if (TheLoop->getNumBackEdges() != 1)
3247 // We must have a single exiting block.
3248 if (!TheLoop->getExitingBlock())
3251 // We need to have a loop header.
3252 DEBUG(dbgs() << "LV: Found a loop: " <<
3253 TheLoop->getHeader()->getName() << '\n');
3255 // Check if we can if-convert non-single-bb loops.
3256 unsigned NumBlocks = TheLoop->getNumBlocks();
3257 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3258 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3262 // ScalarEvolution needs to be able to find the exit count.
3263 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3264 if (ExitCount == SE->getCouldNotCompute()) {
3265 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3269 // Do not loop-vectorize loops with a tiny trip count.
3270 BasicBlock *Latch = TheLoop->getLoopLatch();
3271 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
3272 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
3273 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
3274 "This loop is not worth vectorizing.\n");
3278 // Check if we can vectorize the instructions and CFG in this loop.
3279 if (!canVectorizeInstrs()) {
3280 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3284 // Go over each instruction and look at memory deps.
3285 if (!canVectorizeMemory()) {
3286 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3290 // Collect all of the variables that remain uniform after vectorization.
3291 collectLoopUniforms();
3293 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3294 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3297 // Okay! We can vectorize. At this point we don't have any other mem analysis
3298 // which may limit our maximum vectorization factor, so just return true with
3303 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3304 if (Ty->isPointerTy())
3305 return DL.getIntPtrType(Ty);
3307 // It is possible that char's or short's overflow when we ask for the loop's
3308 // trip count, work around this by changing the type size.
3309 if (Ty->getScalarSizeInBits() < 32)
3310 return Type::getInt32Ty(Ty->getContext());
3315 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3316 Ty0 = convertPointerToIntegerType(DL, Ty0);
3317 Ty1 = convertPointerToIntegerType(DL, Ty1);
3318 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3323 /// \brief Check that the instruction has outside loop users and is not an
3324 /// identified reduction variable.
3325 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3326 SmallPtrSet<Value *, 4> &Reductions) {
3327 // Reduction instructions are allowed to have exit users. All other
3328 // instructions must not have external users.
3329 if (!Reductions.count(Inst))
3330 //Check that all of the users of the loop are inside the BB.
3331 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
3333 Instruction *U = cast<Instruction>(*I);
3334 // This user may be a reduction exit value.
3335 if (!TheLoop->contains(U)) {
3336 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
3343 bool LoopVectorizationLegality::canVectorizeInstrs() {
3344 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3345 BasicBlock *Header = TheLoop->getHeader();
3347 // Look for the attribute signaling the absence of NaNs.
3348 Function &F = *Header->getParent();
3349 if (F.hasFnAttribute("no-nans-fp-math"))
3350 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3351 AttributeSet::FunctionIndex,
3352 "no-nans-fp-math").getValueAsString() == "true";
3354 // For each block in the loop.
3355 for (Loop::block_iterator bb = TheLoop->block_begin(),
3356 be = TheLoop->block_end(); bb != be; ++bb) {
3358 // Scan the instructions in the block and look for hazards.
3359 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3362 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3363 Type *PhiTy = Phi->getType();
3364 // Check that this PHI type is allowed.
3365 if (!PhiTy->isIntegerTy() &&
3366 !PhiTy->isFloatingPointTy() &&
3367 !PhiTy->isPointerTy()) {
3368 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3372 // If this PHINode is not in the header block, then we know that we
3373 // can convert it to select during if-conversion. No need to check if
3374 // the PHIs in this block are induction or reduction variables.
3375 if (*bb != Header) {
3376 // Check that this instruction has no outside users or is an
3377 // identified reduction value with an outside user.
3378 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3383 // We only allow if-converted PHIs with more than two incoming values.
3384 if (Phi->getNumIncomingValues() != 2) {
3385 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3389 // This is the value coming from the preheader.
3390 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3391 // Check if this is an induction variable.
3392 InductionKind IK = isInductionVariable(Phi);
3394 if (IK_NoInduction != IK) {
3395 // Get the widest type.
3397 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3399 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3401 // Int inductions are special because we only allow one IV.
3402 if (IK == IK_IntInduction) {
3403 // Use the phi node with the widest type as induction. Use the last
3404 // one if there are multiple (no good reason for doing this other
3405 // than it is expedient).
3406 if (!Induction || PhiTy == WidestIndTy)
3410 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3411 Inductions[Phi] = InductionInfo(StartValue, IK);
3413 // Until we explicitly handle the case of an induction variable with
3414 // an outside loop user we have to give up vectorizing this loop.
3415 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3421 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3422 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3425 if (AddReductionVar(Phi, RK_IntegerMult)) {
3426 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3429 if (AddReductionVar(Phi, RK_IntegerOr)) {
3430 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3433 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3434 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3437 if (AddReductionVar(Phi, RK_IntegerXor)) {
3438 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3441 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3442 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3445 if (AddReductionVar(Phi, RK_FloatMult)) {
3446 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3449 if (AddReductionVar(Phi, RK_FloatAdd)) {
3450 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3453 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3454 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3459 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3461 }// end of PHI handling
3463 // We still don't handle functions. However, we can ignore dbg intrinsic
3464 // calls and we do handle certain intrinsic and libm functions.
3465 CallInst *CI = dyn_cast<CallInst>(it);
3466 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3467 DEBUG(dbgs() << "LV: Found a call site.\n");
3471 // Check that the instruction return type is vectorizable.
3472 // Also, we can't vectorize extractelement instructions.
3473 if ((!VectorType::isValidElementType(it->getType()) &&
3474 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3475 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3479 // Check that the stored type is vectorizable.
3480 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3481 Type *T = ST->getValueOperand()->getType();
3482 if (!VectorType::isValidElementType(T))
3484 if (EnableMemAccessVersioning)
3485 collectStridedAcccess(ST);
3488 if (EnableMemAccessVersioning)
3489 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3490 collectStridedAcccess(LI);
3492 // Reduction instructions are allowed to have exit users.
3493 // All other instructions must not have external users.
3494 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3502 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3503 if (Inductions.empty())
3510 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3511 /// return the induction operand of the gep pointer.
3512 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3513 const DataLayout *DL, Loop *Lp) {
3514 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3518 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3520 // Check that all of the gep indices are uniform except for our induction
3522 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3523 if (i != InductionOperand &&
3524 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3526 return GEP->getOperand(InductionOperand);
3529 ///\brief Look for a cast use of the passed value.
3530 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3531 Value *UniqueCast = 0;
3532 for (Value::use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end(); UI != UE;
3534 CastInst *CI = dyn_cast<CastInst>(*UI);
3535 if (CI && CI->getType() == Ty) {
3545 ///\brief Get the stride of a pointer access in a loop.
3546 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3547 /// pointer to the Value, or null otherwise.
3548 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3549 const DataLayout *DL, Loop *Lp) {
3550 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3551 if (!PtrTy || PtrTy->isAggregateType())
3554 // Try to remove a gep instruction to make the pointer (actually index at this
3555 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3556 // pointer, otherwise, we are analyzing the index.
3557 Value *OrigPtr = Ptr;
3559 // The size of the pointer access.
3560 int64_t PtrAccessSize = 1;
3562 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3563 const SCEV *V = SE->getSCEV(Ptr);
3567 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3568 V = C->getOperand();
3570 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3574 V = S->getStepRecurrence(*SE);
3578 // Strip off the size of access multiplication if we are still analyzing the
3580 if (OrigPtr == Ptr) {
3581 DL->getTypeAllocSize(PtrTy->getElementType());
3582 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3583 if (M->getOperand(0)->getSCEVType() != scConstant)
3586 const APInt &APStepVal =
3587 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3589 // Huge step value - give up.
3590 if (APStepVal.getBitWidth() > 64)
3593 int64_t StepVal = APStepVal.getSExtValue();
3594 if (PtrAccessSize != StepVal)
3596 V = M->getOperand(1);
3601 Type *StripedOffRecurrenceCast = 0;
3602 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3603 StripedOffRecurrenceCast = C->getType();
3604 V = C->getOperand();
3607 // Look for the loop invariant symbolic value.
3608 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3612 Value *Stride = U->getValue();
3613 if (!Lp->isLoopInvariant(Stride))
3616 // If we have stripped off the recurrence cast we have to make sure that we
3617 // return the value that is used in this loop so that we can replace it later.
3618 if (StripedOffRecurrenceCast)
3619 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3624 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3626 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3627 Ptr = LI->getPointerOperand();
3628 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3629 Ptr = SI->getPointerOperand();
3633 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3637 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3638 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3639 Strides[Ptr] = Stride;
3640 StrideSet.insert(Stride);
3643 void LoopVectorizationLegality::collectLoopUniforms() {
3644 // We now know that the loop is vectorizable!
3645 // Collect variables that will remain uniform after vectorization.
3646 std::vector<Value*> Worklist;
3647 BasicBlock *Latch = TheLoop->getLoopLatch();
3649 // Start with the conditional branch and walk up the block.
3650 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3652 while (Worklist.size()) {
3653 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3654 Worklist.pop_back();
3656 // Look at instructions inside this loop.
3657 // Stop when reaching PHI nodes.
3658 // TODO: we need to follow values all over the loop, not only in this block.
3659 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3662 // This is a known uniform.
3665 // Insert all operands.
3666 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3671 /// \brief Analyses memory accesses in a loop.
3673 /// Checks whether run time pointer checks are needed and builds sets for data
3674 /// dependence checking.
3675 class AccessAnalysis {
3677 /// \brief Read or write access location.
3678 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3679 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3681 /// \brief Set of potential dependent memory accesses.
3682 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3684 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3685 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3686 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3688 /// \brief Register a load and whether it is only read from.
3689 void addLoad(Value *Ptr, bool IsReadOnly) {
3690 Accesses.insert(MemAccessInfo(Ptr, false));
3692 ReadOnlyPtr.insert(Ptr);
3695 /// \brief Register a store.
3696 void addStore(Value *Ptr) {
3697 Accesses.insert(MemAccessInfo(Ptr, true));
3700 /// \brief Check whether we can check the pointers at runtime for
3701 /// non-intersection.
3702 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3703 unsigned &NumComparisons, ScalarEvolution *SE,
3704 Loop *TheLoop, ValueToValueMap &Strides,
3705 bool ShouldCheckStride = false);
3707 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3708 /// and builds sets of dependent accesses.
3709 void buildDependenceSets() {
3710 // Process read-write pointers first.
3711 processMemAccesses(false);
3712 // Next, process read pointers.
3713 processMemAccesses(true);
3716 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3718 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3719 void resetDepChecks() { CheckDeps.clear(); }
3721 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3724 typedef SetVector<MemAccessInfo> PtrAccessSet;
3725 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3727 /// \brief Go over all memory access or only the deferred ones if
3728 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3729 /// and build sets of dependency check candidates.
3730 void processMemAccesses(bool UseDeferred);
3732 /// Set of all accesses.
3733 PtrAccessSet Accesses;
3735 /// Set of access to check after all writes have been processed.
3736 PtrAccessSet DeferredAccesses;
3738 /// Map of pointers to last access encountered.
3739 UnderlyingObjToAccessMap ObjToLastAccess;
3741 /// Set of accesses that need a further dependence check.
3742 MemAccessInfoSet CheckDeps;
3744 /// Set of pointers that are read only.
3745 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3747 /// Set of underlying objects already written to.
3748 SmallPtrSet<Value*, 16> WriteObjects;
3750 const DataLayout *DL;
3752 /// Sets of potentially dependent accesses - members of one set share an
3753 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3754 /// dependence check.
3755 DepCandidates &DepCands;
3757 bool AreAllWritesIdentified;
3758 bool AreAllReadsIdentified;
3759 bool IsRTCheckNeeded;
3762 } // end anonymous namespace
3764 /// \brief Check whether a pointer can participate in a runtime bounds check.
3765 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3767 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3768 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3772 return AR->isAffine();
3775 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3776 /// the address space.
3777 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3778 const Loop *Lp, ValueToValueMap &StridesMap);
3780 bool AccessAnalysis::canCheckPtrAtRT(
3781 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3782 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3783 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3784 // Find pointers with computable bounds. We are going to use this information
3785 // to place a runtime bound check.
3786 unsigned NumReadPtrChecks = 0;
3787 unsigned NumWritePtrChecks = 0;
3788 bool CanDoRT = true;
3790 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3791 // We assign consecutive id to access from different dependence sets.
3792 // Accesses within the same set don't need a runtime check.
3793 unsigned RunningDepId = 1;
3794 DenseMap<Value *, unsigned> DepSetId;
3796 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3798 const MemAccessInfo &Access = *AI;
3799 Value *Ptr = Access.getPointer();
3800 bool IsWrite = Access.getInt();
3802 // Just add write checks if we have both.
3803 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3807 ++NumWritePtrChecks;
3811 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3812 // When we run after a failing dependency check we have to make sure we
3813 // don't have wrapping pointers.
3814 (!ShouldCheckStride ||
3815 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3816 // The id of the dependence set.
3819 if (IsDepCheckNeeded) {
3820 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3821 unsigned &LeaderId = DepSetId[Leader];
3823 LeaderId = RunningDepId++;
3826 // Each access has its own dependence set.
3827 DepId = RunningDepId++;
3829 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3831 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3837 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3838 NumComparisons = 0; // Only one dependence set.
3840 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3841 NumWritePtrChecks - 1));
3844 // If the pointers that we would use for the bounds comparison have different
3845 // address spaces, assume the values aren't directly comparable, so we can't
3846 // use them for the runtime check. We also have to assume they could
3847 // overlap. In the future there should be metadata for whether address spaces
3849 unsigned NumPointers = RtCheck.Pointers.size();
3850 for (unsigned i = 0; i < NumPointers; ++i) {
3851 for (unsigned j = i + 1; j < NumPointers; ++j) {
3852 // Only need to check pointers between two different dependency sets.
3853 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3856 Value *PtrI = RtCheck.Pointers[i];
3857 Value *PtrJ = RtCheck.Pointers[j];
3859 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3860 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3862 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3863 " different address spaces\n");
3872 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3873 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3876 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3877 // We process the set twice: first we process read-write pointers, last we
3878 // process read-only pointers. This allows us to skip dependence tests for
3879 // read-only pointers.
3881 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3882 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3883 const MemAccessInfo &Access = *AI;
3884 Value *Ptr = Access.getPointer();
3885 bool IsWrite = Access.getInt();
3887 DepCands.insert(Access);
3889 // Memorize read-only pointers for later processing and skip them in the
3890 // first round (they need to be checked after we have seen all write
3891 // pointers). Note: we also mark pointer that are not consecutive as
3892 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3893 // second check for "!IsWrite".
3894 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3895 if (!UseDeferred && IsReadOnlyPtr) {
3896 DeferredAccesses.insert(Access);
3900 bool NeedDepCheck = false;
3901 // Check whether there is the possibility of dependency because of
3902 // underlying objects being the same.
3903 typedef SmallVector<Value*, 16> ValueVector;
3904 ValueVector TempObjects;
3905 GetUnderlyingObjects(Ptr, TempObjects, DL);
3906 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3908 Value *UnderlyingObj = *UI;
3910 // If this is a write then it needs to be an identified object. If this a
3911 // read and all writes (so far) are identified function scope objects we
3912 // don't need an identified underlying object but only an Argument (the
3913 // next write is going to invalidate this assumption if it is
3915 // This is a micro-optimization for the case where all writes are
3916 // identified and we have one argument pointer.
3917 // Otherwise, we do need a runtime check.
3918 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3919 (!IsWrite && (!AreAllWritesIdentified ||
3920 !isa<Argument>(UnderlyingObj)) &&
3921 !isIdentifiedObject(UnderlyingObj))) {
3922 DEBUG(dbgs() << "LV: Found an unidentified " <<
3923 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3925 IsRTCheckNeeded = (IsRTCheckNeeded ||
3926 !isIdentifiedObject(UnderlyingObj) ||
3927 !AreAllReadsIdentified);
3930 AreAllWritesIdentified = false;
3932 AreAllReadsIdentified = false;
3935 // If this is a write - check other reads and writes for conflicts. If
3936 // this is a read only check other writes for conflicts (but only if there
3937 // is no other write to the ptr - this is an optimization to catch "a[i] =
3938 // a[i] + " without having to do a dependence check).
3939 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3940 NeedDepCheck = true;
3943 WriteObjects.insert(UnderlyingObj);
3945 // Create sets of pointers connected by shared underlying objects.
3946 UnderlyingObjToAccessMap::iterator Prev =
3947 ObjToLastAccess.find(UnderlyingObj);
3948 if (Prev != ObjToLastAccess.end())
3949 DepCands.unionSets(Access, Prev->second);
3951 ObjToLastAccess[UnderlyingObj] = Access;
3955 CheckDeps.insert(Access);
3960 /// \brief Checks memory dependences among accesses to the same underlying
3961 /// object to determine whether there vectorization is legal or not (and at
3962 /// which vectorization factor).
3964 /// This class works under the assumption that we already checked that memory
3965 /// locations with different underlying pointers are "must-not alias".
3966 /// We use the ScalarEvolution framework to symbolically evalutate access
3967 /// functions pairs. Since we currently don't restructure the loop we can rely
3968 /// on the program order of memory accesses to determine their safety.
3969 /// At the moment we will only deem accesses as safe for:
3970 /// * A negative constant distance assuming program order.
3972 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3973 /// a[i] = tmp; y = a[i];
3975 /// The latter case is safe because later checks guarantuee that there can't
3976 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3977 /// the same variable: a header phi can only be an induction or a reduction, a
3978 /// reduction can't have a memory sink, an induction can't have a memory
3979 /// source). This is important and must not be violated (or we have to
3980 /// resort to checking for cycles through memory).
3982 /// * A positive constant distance assuming program order that is bigger
3983 /// than the biggest memory access.
3985 /// tmp = a[i] OR b[i] = x
3986 /// a[i+2] = tmp y = b[i+2];
3988 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3990 /// * Zero distances and all accesses have the same size.
3992 class MemoryDepChecker {
3994 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3995 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3997 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
3998 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3999 ShouldRetryWithRuntimeCheck(false) {}
4001 /// \brief Register the location (instructions are given increasing numbers)
4002 /// of a write access.
4003 void addAccess(StoreInst *SI) {
4004 Value *Ptr = SI->getPointerOperand();
4005 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4006 InstMap.push_back(SI);
4010 /// \brief Register the location (instructions are given increasing numbers)
4011 /// of a write access.
4012 void addAccess(LoadInst *LI) {
4013 Value *Ptr = LI->getPointerOperand();
4014 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4015 InstMap.push_back(LI);
4019 /// \brief Check whether the dependencies between the accesses are safe.
4021 /// Only checks sets with elements in \p CheckDeps.
4022 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4023 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4025 /// \brief The maximum number of bytes of a vector register we can vectorize
4026 /// the accesses safely with.
4027 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4029 /// \brief In same cases when the dependency check fails we can still
4030 /// vectorize the loop with a dynamic array access check.
4031 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4034 ScalarEvolution *SE;
4035 const DataLayout *DL;
4036 const Loop *InnermostLoop;
4038 /// \brief Maps access locations (ptr, read/write) to program order.
4039 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4041 /// \brief Memory access instructions in program order.
4042 SmallVector<Instruction *, 16> InstMap;
4044 /// \brief The program order index to be used for the next instruction.
4047 // We can access this many bytes in parallel safely.
4048 unsigned MaxSafeDepDistBytes;
4050 /// \brief If we see a non-constant dependence distance we can still try to
4051 /// vectorize this loop with runtime checks.
4052 bool ShouldRetryWithRuntimeCheck;
4054 /// \brief Check whether there is a plausible dependence between the two
4057 /// Access \p A must happen before \p B in program order. The two indices
4058 /// identify the index into the program order map.
4060 /// This function checks whether there is a plausible dependence (or the
4061 /// absence of such can't be proved) between the two accesses. If there is a
4062 /// plausible dependence but the dependence distance is bigger than one
4063 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4064 /// distance is smaller than any other distance encountered so far).
4065 /// Otherwise, this function returns true signaling a possible dependence.
4066 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4067 const MemAccessInfo &B, unsigned BIdx,
4068 ValueToValueMap &Strides);
4070 /// \brief Check whether the data dependence could prevent store-load
4072 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4075 } // end anonymous namespace
4077 static bool isInBoundsGep(Value *Ptr) {
4078 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4079 return GEP->isInBounds();
4083 /// \brief Check whether the access through \p Ptr has a constant stride.
4084 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4085 const Loop *Lp, ValueToValueMap &StridesMap) {
4086 const Type *Ty = Ptr->getType();
4087 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4089 // Make sure that the pointer does not point to aggregate types.
4090 const PointerType *PtrTy = cast<PointerType>(Ty);
4091 if (PtrTy->getElementType()->isAggregateType()) {
4092 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4097 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4099 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4101 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4102 << *Ptr << " SCEV: " << *PtrScev << "\n");
4106 // The accesss function must stride over the innermost loop.
4107 if (Lp != AR->getLoop()) {
4108 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4109 *Ptr << " SCEV: " << *PtrScev << "\n");
4112 // The address calculation must not wrap. Otherwise, a dependence could be
4114 // An inbounds getelementptr that is a AddRec with a unit stride
4115 // cannot wrap per definition. The unit stride requirement is checked later.
4116 // An getelementptr without an inbounds attribute and unit stride would have
4117 // to access the pointer value "0" which is undefined behavior in address
4118 // space 0, therefore we can also vectorize this case.
4119 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4120 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4121 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4122 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4123 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4124 << *Ptr << " SCEV: " << *PtrScev << "\n");
4128 // Check the step is constant.
4129 const SCEV *Step = AR->getStepRecurrence(*SE);
4131 // Calculate the pointer stride and check if it is consecutive.
4132 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4134 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4135 " SCEV: " << *PtrScev << "\n");
4139 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4140 const APInt &APStepVal = C->getValue()->getValue();
4142 // Huge step value - give up.
4143 if (APStepVal.getBitWidth() > 64)
4146 int64_t StepVal = APStepVal.getSExtValue();
4149 int64_t Stride = StepVal / Size;
4150 int64_t Rem = StepVal % Size;
4154 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4155 // know we can't "wrap around the address space". In case of address space
4156 // zero we know that this won't happen without triggering undefined behavior.
4157 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4158 Stride != 1 && Stride != -1)
4164 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4165 unsigned TypeByteSize) {
4166 // If loads occur at a distance that is not a multiple of a feasible vector
4167 // factor store-load forwarding does not take place.
4168 // Positive dependences might cause troubles because vectorizing them might
4169 // prevent store-load forwarding making vectorized code run a lot slower.
4170 // a[i] = a[i-3] ^ a[i-8];
4171 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4172 // hence on your typical architecture store-load forwarding does not take
4173 // place. Vectorizing in such cases does not make sense.
4174 // Store-load forwarding distance.
4175 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4176 // Maximum vector factor.
4177 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4178 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4179 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4181 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4183 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4184 MaxVFWithoutSLForwardIssues = (vf >>=1);
4189 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4190 DEBUG(dbgs() << "LV: Distance " << Distance <<
4191 " that could cause a store-load forwarding conflict\n");
4195 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4196 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4197 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4201 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4202 const MemAccessInfo &B, unsigned BIdx,
4203 ValueToValueMap &Strides) {
4204 assert (AIdx < BIdx && "Must pass arguments in program order");
4206 Value *APtr = A.getPointer();
4207 Value *BPtr = B.getPointer();
4208 bool AIsWrite = A.getInt();
4209 bool BIsWrite = B.getInt();
4211 // Two reads are independent.
4212 if (!AIsWrite && !BIsWrite)
4215 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4216 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4218 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4219 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4221 const SCEV *Src = AScev;
4222 const SCEV *Sink = BScev;
4224 // If the induction step is negative we have to invert source and sink of the
4226 if (StrideAPtr < 0) {
4229 std::swap(APtr, BPtr);
4230 std::swap(Src, Sink);
4231 std::swap(AIsWrite, BIsWrite);
4232 std::swap(AIdx, BIdx);
4233 std::swap(StrideAPtr, StrideBPtr);
4236 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4238 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4239 << "(Induction step: " << StrideAPtr << ")\n");
4240 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4241 << *InstMap[BIdx] << ": " << *Dist << "\n");
4243 // Need consecutive accesses. We don't want to vectorize
4244 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4245 // the address space.
4246 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4247 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4251 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4253 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4254 ShouldRetryWithRuntimeCheck = true;
4258 Type *ATy = APtr->getType()->getPointerElementType();
4259 Type *BTy = BPtr->getType()->getPointerElementType();
4260 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4262 // Negative distances are not plausible dependencies.
4263 const APInt &Val = C->getValue()->getValue();
4264 if (Val.isNegative()) {
4265 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4266 if (IsTrueDataDependence &&
4267 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4271 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4275 // Write to the same location with the same size.
4276 // Could be improved to assert type sizes are the same (i32 == float, etc).
4280 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4284 assert(Val.isStrictlyPositive() && "Expect a positive value");
4286 // Positive distance bigger than max vectorization factor.
4289 "LV: ReadWrite-Write positive dependency with different types\n");
4293 unsigned Distance = (unsigned) Val.getZExtValue();
4295 // Bail out early if passed-in parameters make vectorization not feasible.
4296 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4297 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4299 // The distance must be bigger than the size needed for a vectorized version
4300 // of the operation and the size of the vectorized operation must not be
4301 // bigger than the currrent maximum size.
4302 if (Distance < 2*TypeByteSize ||
4303 2*TypeByteSize > MaxSafeDepDistBytes ||
4304 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4305 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4306 << Val.getSExtValue() << '\n');
4310 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4311 Distance : MaxSafeDepDistBytes;
4313 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4314 if (IsTrueDataDependence &&
4315 couldPreventStoreLoadForward(Distance, TypeByteSize))
4318 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4319 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4324 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4325 MemAccessInfoSet &CheckDeps,
4326 ValueToValueMap &Strides) {
4328 MaxSafeDepDistBytes = -1U;
4329 while (!CheckDeps.empty()) {
4330 MemAccessInfo CurAccess = *CheckDeps.begin();
4332 // Get the relevant memory access set.
4333 EquivalenceClasses<MemAccessInfo>::iterator I =
4334 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4336 // Check accesses within this set.
4337 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4338 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4340 // Check every access pair.
4342 CheckDeps.erase(*AI);
4343 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4345 // Check every accessing instruction pair in program order.
4346 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4347 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4348 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4349 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4350 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4352 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4363 bool LoopVectorizationLegality::canVectorizeMemory() {
4365 typedef SmallVector<Value*, 16> ValueVector;
4366 typedef SmallPtrSet<Value*, 16> ValueSet;
4368 // Holds the Load and Store *instructions*.
4372 // Holds all the different accesses in the loop.
4373 unsigned NumReads = 0;
4374 unsigned NumReadWrites = 0;
4376 PtrRtCheck.Pointers.clear();
4377 PtrRtCheck.Need = false;
4379 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4380 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4383 for (Loop::block_iterator bb = TheLoop->block_begin(),
4384 be = TheLoop->block_end(); bb != be; ++bb) {
4386 // Scan the BB and collect legal loads and stores.
4387 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4390 // If this is a load, save it. If this instruction can read from memory
4391 // but is not a load, then we quit. Notice that we don't handle function
4392 // calls that read or write.
4393 if (it->mayReadFromMemory()) {
4394 // Many math library functions read the rounding mode. We will only
4395 // vectorize a loop if it contains known function calls that don't set
4396 // the flag. Therefore, it is safe to ignore this read from memory.
4397 CallInst *Call = dyn_cast<CallInst>(it);
4398 if (Call && getIntrinsicIDForCall(Call, TLI))
4401 LoadInst *Ld = dyn_cast<LoadInst>(it);
4402 if (!Ld) return false;
4403 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4404 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4408 Loads.push_back(Ld);
4409 DepChecker.addAccess(Ld);
4413 // Save 'store' instructions. Abort if other instructions write to memory.
4414 if (it->mayWriteToMemory()) {
4415 StoreInst *St = dyn_cast<StoreInst>(it);
4416 if (!St) return false;
4417 if (!St->isSimple() && !IsAnnotatedParallel) {
4418 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4422 Stores.push_back(St);
4423 DepChecker.addAccess(St);
4428 // Now we have two lists that hold the loads and the stores.
4429 // Next, we find the pointers that they use.
4431 // Check if we see any stores. If there are no stores, then we don't
4432 // care if the pointers are *restrict*.
4433 if (!Stores.size()) {
4434 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4438 AccessAnalysis::DepCandidates DependentAccesses;
4439 AccessAnalysis Accesses(DL, DependentAccesses);
4441 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4442 // multiple times on the same object. If the ptr is accessed twice, once
4443 // for read and once for write, it will only appear once (on the write
4444 // list). This is okay, since we are going to check for conflicts between
4445 // writes and between reads and writes, but not between reads and reads.
4448 ValueVector::iterator I, IE;
4449 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4450 StoreInst *ST = cast<StoreInst>(*I);
4451 Value* Ptr = ST->getPointerOperand();
4453 if (isUniform(Ptr)) {
4454 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4458 // If we did *not* see this pointer before, insert it to the read-write
4459 // list. At this phase it is only a 'write' list.
4460 if (Seen.insert(Ptr)) {
4462 Accesses.addStore(Ptr);
4466 if (IsAnnotatedParallel) {
4468 << "LV: A loop annotated parallel, ignore memory dependency "
4473 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4474 LoadInst *LD = cast<LoadInst>(*I);
4475 Value* Ptr = LD->getPointerOperand();
4476 // If we did *not* see this pointer before, insert it to the
4477 // read list. If we *did* see it before, then it is already in
4478 // the read-write list. This allows us to vectorize expressions
4479 // such as A[i] += x; Because the address of A[i] is a read-write
4480 // pointer. This only works if the index of A[i] is consecutive.
4481 // If the address of i is unknown (for example A[B[i]]) then we may
4482 // read a few words, modify, and write a few words, and some of the
4483 // words may be written to the same address.
4484 bool IsReadOnlyPtr = false;
4485 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4487 IsReadOnlyPtr = true;
4489 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4492 // If we write (or read-write) to a single destination and there are no
4493 // other reads in this loop then is it safe to vectorize.
4494 if (NumReadWrites == 1 && NumReads == 0) {
4495 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4499 // Build dependence sets and check whether we need a runtime pointer bounds
4501 Accesses.buildDependenceSets();
4502 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4504 // Find pointers with computable bounds. We are going to use this information
4505 // to place a runtime bound check.
4506 unsigned NumComparisons = 0;
4507 bool CanDoRT = false;
4509 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4512 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4513 " pointer comparisons.\n");
4515 // If we only have one set of dependences to check pointers among we don't
4516 // need a runtime check.
4517 if (NumComparisons == 0 && NeedRTCheck)
4518 NeedRTCheck = false;
4520 // Check that we did not collect too many pointers or found an unsizeable
4522 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4528 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4531 if (NeedRTCheck && !CanDoRT) {
4532 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4533 "the array bounds.\n");
4538 PtrRtCheck.Need = NeedRTCheck;
4540 bool CanVecMem = true;
4541 if (Accesses.isDependencyCheckNeeded()) {
4542 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4543 CanVecMem = DepChecker.areDepsSafe(
4544 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4545 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4547 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4548 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4551 // Clear the dependency checks. We assume they are not needed.
4552 Accesses.resetDepChecks();
4555 PtrRtCheck.Need = true;
4557 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4558 TheLoop, Strides, true);
4559 // Check that we did not collect too many pointers or found an unsizeable
4561 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4562 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4571 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4572 " need a runtime memory check.\n");
4577 static bool hasMultipleUsesOf(Instruction *I,
4578 SmallPtrSet<Instruction *, 8> &Insts) {
4579 unsigned NumUses = 0;
4580 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4581 if (Insts.count(dyn_cast<Instruction>(*Use)))
4590 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4591 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4592 if (!Set.count(dyn_cast<Instruction>(*Use)))
4597 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4598 ReductionKind Kind) {
4599 if (Phi->getNumIncomingValues() != 2)
4602 // Reduction variables are only found in the loop header block.
4603 if (Phi->getParent() != TheLoop->getHeader())
4606 // Obtain the reduction start value from the value that comes from the loop
4608 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4610 // ExitInstruction is the single value which is used outside the loop.
4611 // We only allow for a single reduction value to be used outside the loop.
4612 // This includes users of the reduction, variables (which form a cycle
4613 // which ends in the phi node).
4614 Instruction *ExitInstruction = 0;
4615 // Indicates that we found a reduction operation in our scan.
4616 bool FoundReduxOp = false;
4618 // We start with the PHI node and scan for all of the users of this
4619 // instruction. All users must be instructions that can be used as reduction
4620 // variables (such as ADD). We must have a single out-of-block user. The cycle
4621 // must include the original PHI.
4622 bool FoundStartPHI = false;
4624 // To recognize min/max patterns formed by a icmp select sequence, we store
4625 // the number of instruction we saw from the recognized min/max pattern,
4626 // to make sure we only see exactly the two instructions.
4627 unsigned NumCmpSelectPatternInst = 0;
4628 ReductionInstDesc ReduxDesc(false, 0);
4630 SmallPtrSet<Instruction *, 8> VisitedInsts;
4631 SmallVector<Instruction *, 8> Worklist;
4632 Worklist.push_back(Phi);
4633 VisitedInsts.insert(Phi);
4635 // A value in the reduction can be used:
4636 // - By the reduction:
4637 // - Reduction operation:
4638 // - One use of reduction value (safe).
4639 // - Multiple use of reduction value (not safe).
4641 // - All uses of the PHI must be the reduction (safe).
4642 // - Otherwise, not safe.
4643 // - By one instruction outside of the loop (safe).
4644 // - By further instructions outside of the loop (not safe).
4645 // - By an instruction that is not part of the reduction (not safe).
4647 // * An instruction type other than PHI or the reduction operation.
4648 // * A PHI in the header other than the initial PHI.
4649 while (!Worklist.empty()) {
4650 Instruction *Cur = Worklist.back();
4651 Worklist.pop_back();
4654 // If the instruction has no users then this is a broken chain and can't be
4655 // a reduction variable.
4656 if (Cur->use_empty())
4659 bool IsAPhi = isa<PHINode>(Cur);
4661 // A header PHI use other than the original PHI.
4662 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4665 // Reductions of instructions such as Div, and Sub is only possible if the
4666 // LHS is the reduction variable.
4667 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4668 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4669 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4672 // Any reduction instruction must be of one of the allowed kinds.
4673 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4674 if (!ReduxDesc.IsReduction)
4677 // A reduction operation must only have one use of the reduction value.
4678 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4679 hasMultipleUsesOf(Cur, VisitedInsts))
4682 // All inputs to a PHI node must be a reduction value.
4683 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4686 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4687 isa<SelectInst>(Cur)))
4688 ++NumCmpSelectPatternInst;
4689 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4690 isa<SelectInst>(Cur)))
4691 ++NumCmpSelectPatternInst;
4693 // Check whether we found a reduction operator.
4694 FoundReduxOp |= !IsAPhi;
4696 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4697 // onto the stack. This way we are going to have seen all inputs to PHI
4698 // nodes once we get to them.
4699 SmallVector<Instruction *, 8> NonPHIs;
4700 SmallVector<Instruction *, 8> PHIs;
4701 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4703 Instruction *Usr = cast<Instruction>(*UI);
4705 // Check if we found the exit user.
4706 BasicBlock *Parent = Usr->getParent();
4707 if (!TheLoop->contains(Parent)) {
4708 // Exit if you find multiple outside users or if the header phi node is
4709 // being used. In this case the user uses the value of the previous
4710 // iteration, in which case we would loose "VF-1" iterations of the
4711 // reduction operation if we vectorize.
4712 if (ExitInstruction != 0 || Cur == Phi)
4715 // The instruction used by an outside user must be the last instruction
4716 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4717 // operations on the value.
4718 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4721 ExitInstruction = Cur;
4725 // Process instructions only once (termination). Each reduction cycle
4726 // value must only be used once, except by phi nodes and min/max
4727 // reductions which are represented as a cmp followed by a select.
4728 ReductionInstDesc IgnoredVal(false, 0);
4729 if (VisitedInsts.insert(Usr)) {
4730 if (isa<PHINode>(Usr))
4731 PHIs.push_back(Usr);
4733 NonPHIs.push_back(Usr);
4734 } else if (!isa<PHINode>(Usr) &&
4735 ((!isa<FCmpInst>(Usr) &&
4736 !isa<ICmpInst>(Usr) &&
4737 !isa<SelectInst>(Usr)) ||
4738 !isMinMaxSelectCmpPattern(Usr, IgnoredVal).IsReduction))
4741 // Remember that we completed the cycle.
4743 FoundStartPHI = true;
4745 Worklist.append(PHIs.begin(), PHIs.end());
4746 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4749 // This means we have seen one but not the other instruction of the
4750 // pattern or more than just a select and cmp.
4751 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4752 NumCmpSelectPatternInst != 2)
4755 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4758 // We found a reduction var if we have reached the original phi node and we
4759 // only have a single instruction with out-of-loop users.
4761 // This instruction is allowed to have out-of-loop users.
4762 AllowedExit.insert(ExitInstruction);
4764 // Save the description of this reduction variable.
4765 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4766 ReduxDesc.MinMaxKind);
4767 Reductions[Phi] = RD;
4768 // We've ended the cycle. This is a reduction variable if we have an
4769 // outside user and it has a binary op.
4774 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4775 /// pattern corresponding to a min(X, Y) or max(X, Y).
4776 LoopVectorizationLegality::ReductionInstDesc
4777 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4778 ReductionInstDesc &Prev) {
4780 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4781 "Expect a select instruction");
4782 Instruction *Cmp = 0;
4783 SelectInst *Select = 0;
4785 // We must handle the select(cmp()) as a single instruction. Advance to the
4787 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4788 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4789 return ReductionInstDesc(false, I);
4790 return ReductionInstDesc(Select, Prev.MinMaxKind);
4793 // Only handle single use cases for now.
4794 if (!(Select = dyn_cast<SelectInst>(I)))
4795 return ReductionInstDesc(false, I);
4796 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4797 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4798 return ReductionInstDesc(false, I);
4799 if (!Cmp->hasOneUse())
4800 return ReductionInstDesc(false, I);
4805 // Look for a min/max pattern.
4806 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4807 return ReductionInstDesc(Select, MRK_UIntMin);
4808 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4809 return ReductionInstDesc(Select, MRK_UIntMax);
4810 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4811 return ReductionInstDesc(Select, MRK_SIntMax);
4812 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4813 return ReductionInstDesc(Select, MRK_SIntMin);
4814 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4815 return ReductionInstDesc(Select, MRK_FloatMin);
4816 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4817 return ReductionInstDesc(Select, MRK_FloatMax);
4818 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4819 return ReductionInstDesc(Select, MRK_FloatMin);
4820 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4821 return ReductionInstDesc(Select, MRK_FloatMax);
4823 return ReductionInstDesc(false, I);
4826 LoopVectorizationLegality::ReductionInstDesc
4827 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4829 ReductionInstDesc &Prev) {
4830 bool FP = I->getType()->isFloatingPointTy();
4831 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4832 switch (I->getOpcode()) {
4834 return ReductionInstDesc(false, I);
4835 case Instruction::PHI:
4836 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4837 Kind != RK_FloatMinMax))
4838 return ReductionInstDesc(false, I);
4839 return ReductionInstDesc(I, Prev.MinMaxKind);
4840 case Instruction::Sub:
4841 case Instruction::Add:
4842 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4843 case Instruction::Mul:
4844 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4845 case Instruction::And:
4846 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4847 case Instruction::Or:
4848 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4849 case Instruction::Xor:
4850 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4851 case Instruction::FMul:
4852 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4853 case Instruction::FAdd:
4854 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4855 case Instruction::FCmp:
4856 case Instruction::ICmp:
4857 case Instruction::Select:
4858 if (Kind != RK_IntegerMinMax &&
4859 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4860 return ReductionInstDesc(false, I);
4861 return isMinMaxSelectCmpPattern(I, Prev);
4865 LoopVectorizationLegality::InductionKind
4866 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4867 Type *PhiTy = Phi->getType();
4868 // We only handle integer and pointer inductions variables.
4869 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4870 return IK_NoInduction;
4872 // Check that the PHI is consecutive.
4873 const SCEV *PhiScev = SE->getSCEV(Phi);
4874 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4876 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4877 return IK_NoInduction;
4879 const SCEV *Step = AR->getStepRecurrence(*SE);
4881 // Integer inductions need to have a stride of one.
4882 if (PhiTy->isIntegerTy()) {
4884 return IK_IntInduction;
4885 if (Step->isAllOnesValue())
4886 return IK_ReverseIntInduction;
4887 return IK_NoInduction;
4890 // Calculate the pointer stride and check if it is consecutive.
4891 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4893 return IK_NoInduction;
4895 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4896 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4897 if (C->getValue()->equalsInt(Size))
4898 return IK_PtrInduction;
4899 else if (C->getValue()->equalsInt(0 - Size))
4900 return IK_ReversePtrInduction;
4902 return IK_NoInduction;
4905 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4906 Value *In0 = const_cast<Value*>(V);
4907 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4911 return Inductions.count(PN);
4914 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4915 assert(TheLoop->contains(BB) && "Unknown block used");
4917 // Blocks that do not dominate the latch need predication.
4918 BasicBlock* Latch = TheLoop->getLoopLatch();
4919 return !DT->dominates(BB, Latch);
4922 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4923 SmallPtrSet<Value *, 8>& SafePtrs) {
4924 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4925 // We might be able to hoist the load.
4926 if (it->mayReadFromMemory()) {
4927 LoadInst *LI = dyn_cast<LoadInst>(it);
4928 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4932 // We don't predicate stores at the moment.
4933 if (it->mayWriteToMemory()) {
4934 StoreInst *SI = dyn_cast<StoreInst>(it);
4935 // We only support predication of stores in basic blocks with one
4937 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
4938 !SafePtrs.count(SI->getPointerOperand()) ||
4939 !SI->getParent()->getSinglePredecessor())
4945 // Check that we don't have a constant expression that can trap as operand.
4946 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4948 if (Constant *C = dyn_cast<Constant>(*OI))
4953 // The instructions below can trap.
4954 switch (it->getOpcode()) {
4956 case Instruction::UDiv:
4957 case Instruction::SDiv:
4958 case Instruction::URem:
4959 case Instruction::SRem:
4967 LoopVectorizationCostModel::VectorizationFactor
4968 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4970 // Width 1 means no vectorize
4971 VectorizationFactor Factor = { 1U, 0U };
4972 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4973 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4977 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
4978 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4982 // Find the trip count.
4983 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4984 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4986 unsigned WidestType = getWidestType();
4987 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4988 unsigned MaxSafeDepDist = -1U;
4989 if (Legal->getMaxSafeDepDistBytes() != -1U)
4990 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4991 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4992 WidestRegister : MaxSafeDepDist);
4993 unsigned MaxVectorSize = WidestRegister / WidestType;
4994 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4995 DEBUG(dbgs() << "LV: The Widest register is: "
4996 << WidestRegister << " bits.\n");
4998 if (MaxVectorSize == 0) {
4999 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5003 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5004 " into one vector!");
5006 unsigned VF = MaxVectorSize;
5008 // If we optimize the program for size, avoid creating the tail loop.
5010 // If we are unable to calculate the trip count then don't try to vectorize.
5012 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5016 // Find the maximum SIMD width that can fit within the trip count.
5017 VF = TC % MaxVectorSize;
5022 // If the trip count that we found modulo the vectorization factor is not
5023 // zero then we require a tail.
5025 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5031 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5032 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5034 Factor.Width = UserVF;
5038 float Cost = expectedCost(1);
5040 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
5041 for (unsigned i=2; i <= VF; i*=2) {
5042 // Notice that the vector loop needs to be executed less times, so
5043 // we need to divide the cost of the vector loops by the width of
5044 // the vector elements.
5045 float VectorCost = expectedCost(i) / (float)i;
5046 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5047 (int)VectorCost << ".\n");
5048 if (VectorCost < Cost) {
5054 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
5055 Factor.Width = Width;
5056 Factor.Cost = Width * Cost;
5060 unsigned LoopVectorizationCostModel::getWidestType() {
5061 unsigned MaxWidth = 8;
5064 for (Loop::block_iterator bb = TheLoop->block_begin(),
5065 be = TheLoop->block_end(); bb != be; ++bb) {
5066 BasicBlock *BB = *bb;
5068 // For each instruction in the loop.
5069 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5070 Type *T = it->getType();
5072 // Only examine Loads, Stores and PHINodes.
5073 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5076 // Examine PHI nodes that are reduction variables.
5077 if (PHINode *PN = dyn_cast<PHINode>(it))
5078 if (!Legal->getReductionVars()->count(PN))
5081 // Examine the stored values.
5082 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5083 T = ST->getValueOperand()->getType();
5085 // Ignore loaded pointer types and stored pointer types that are not
5086 // consecutive. However, we do want to take consecutive stores/loads of
5087 // pointer vectors into account.
5088 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5091 MaxWidth = std::max(MaxWidth,
5092 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5100 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5103 unsigned LoopCost) {
5105 // -- The unroll heuristics --
5106 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5107 // There are many micro-architectural considerations that we can't predict
5108 // at this level. For example frontend pressure (on decode or fetch) due to
5109 // code size, or the number and capabilities of the execution ports.
5111 // We use the following heuristics to select the unroll factor:
5112 // 1. If the code has reductions the we unroll in order to break the cross
5113 // iteration dependency.
5114 // 2. If the loop is really small then we unroll in order to reduce the loop
5116 // 3. We don't unroll if we think that we will spill registers to memory due
5117 // to the increased register pressure.
5119 // Use the user preference, unless 'auto' is selected.
5123 // When we optimize for size we don't unroll.
5127 // We used the distance for the unroll factor.
5128 if (Legal->getMaxSafeDepDistBytes() != -1U)
5131 // Do not unroll loops with a relatively small trip count.
5132 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5133 TheLoop->getLoopLatch());
5134 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5137 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5138 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5142 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5143 TargetNumRegisters = ForceTargetNumScalarRegs;
5145 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5146 TargetNumRegisters = ForceTargetNumVectorRegs;
5149 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5150 // We divide by these constants so assume that we have at least one
5151 // instruction that uses at least one register.
5152 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5153 R.NumInstructions = std::max(R.NumInstructions, 1U);
5155 // We calculate the unroll factor using the following formula.
5156 // Subtract the number of loop invariants from the number of available
5157 // registers. These registers are used by all of the unrolled instances.
5158 // Next, divide the remaining registers by the number of registers that is
5159 // required by the loop, in order to estimate how many parallel instances
5160 // fit without causing spills. All of this is rounded down if necessary to be
5161 // a power of two. We want power of two unroll factors to simplify any
5162 // addressing operations or alignment considerations.
5163 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5166 // Don't count the induction variable as unrolled.
5167 if (EnableIndVarRegisterHeur)
5168 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5169 std::max(1U, (R.MaxLocalUsers - 1)));
5171 // Clamp the unroll factor ranges to reasonable factors.
5172 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5174 // Check if the user has overridden the unroll max.
5176 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5177 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5179 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5180 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5183 // If we did not calculate the cost for VF (because the user selected the VF)
5184 // then we calculate the cost of VF here.
5186 LoopCost = expectedCost(VF);
5188 // Clamp the calculated UF to be between the 1 and the max unroll factor
5189 // that the target allows.
5190 if (UF > MaxUnrollSize)
5195 // Unroll if we vectorized this loop and there is a reduction that could
5196 // benefit from unrolling.
5197 if (VF > 1 && Legal->getReductionVars()->size()) {
5198 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5202 // Note that if we've already vectorized the loop we will have done the
5203 // runtime check and so unrolling won't require further checks.
5204 bool UnrollingRequiresRuntimePointerCheck =
5205 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5207 // We want to unroll small loops in order to reduce the loop overhead and
5208 // potentially expose ILP opportunities.
5209 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5210 if (!UnrollingRequiresRuntimePointerCheck &&
5211 LoopCost < SmallLoopCost) {
5212 // We assume that the cost overhead is 1 and we use the cost model
5213 // to estimate the cost of the loop and unroll until the cost of the
5214 // loop overhead is about 5% of the cost of the loop.
5215 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5217 // Unroll until store/load ports (estimated by max unroll factor) are
5219 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5220 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5222 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5223 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5224 return std::max(StoresUF, LoadsUF);
5227 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5231 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5235 LoopVectorizationCostModel::RegisterUsage
5236 LoopVectorizationCostModel::calculateRegisterUsage() {
5237 // This function calculates the register usage by measuring the highest number
5238 // of values that are alive at a single location. Obviously, this is a very
5239 // rough estimation. We scan the loop in a topological order in order and
5240 // assign a number to each instruction. We use RPO to ensure that defs are
5241 // met before their users. We assume that each instruction that has in-loop
5242 // users starts an interval. We record every time that an in-loop value is
5243 // used, so we have a list of the first and last occurrences of each
5244 // instruction. Next, we transpose this data structure into a multi map that
5245 // holds the list of intervals that *end* at a specific location. This multi
5246 // map allows us to perform a linear search. We scan the instructions linearly
5247 // and record each time that a new interval starts, by placing it in a set.
5248 // If we find this value in the multi-map then we remove it from the set.
5249 // The max register usage is the maximum size of the set.
5250 // We also search for instructions that are defined outside the loop, but are
5251 // used inside the loop. We need this number separately from the max-interval
5252 // usage number because when we unroll, loop-invariant values do not take
5254 LoopBlocksDFS DFS(TheLoop);
5258 R.NumInstructions = 0;
5260 // Each 'key' in the map opens a new interval. The values
5261 // of the map are the index of the 'last seen' usage of the
5262 // instruction that is the key.
5263 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5264 // Maps instruction to its index.
5265 DenseMap<unsigned, Instruction*> IdxToInstr;
5266 // Marks the end of each interval.
5267 IntervalMap EndPoint;
5268 // Saves the list of instruction indices that are used in the loop.
5269 SmallSet<Instruction*, 8> Ends;
5270 // Saves the list of values that are used in the loop but are
5271 // defined outside the loop, such as arguments and constants.
5272 SmallPtrSet<Value*, 8> LoopInvariants;
5275 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5276 be = DFS.endRPO(); bb != be; ++bb) {
5277 R.NumInstructions += (*bb)->size();
5278 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5280 Instruction *I = it;
5281 IdxToInstr[Index++] = I;
5283 // Save the end location of each USE.
5284 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5285 Value *U = I->getOperand(i);
5286 Instruction *Instr = dyn_cast<Instruction>(U);
5288 // Ignore non-instruction values such as arguments, constants, etc.
5289 if (!Instr) continue;
5291 // If this instruction is outside the loop then record it and continue.
5292 if (!TheLoop->contains(Instr)) {
5293 LoopInvariants.insert(Instr);
5297 // Overwrite previous end points.
5298 EndPoint[Instr] = Index;
5304 // Saves the list of intervals that end with the index in 'key'.
5305 typedef SmallVector<Instruction*, 2> InstrList;
5306 DenseMap<unsigned, InstrList> TransposeEnds;
5308 // Transpose the EndPoints to a list of values that end at each index.
5309 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5311 TransposeEnds[it->second].push_back(it->first);
5313 SmallSet<Instruction*, 8> OpenIntervals;
5314 unsigned MaxUsage = 0;
5317 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5318 for (unsigned int i = 0; i < Index; ++i) {
5319 Instruction *I = IdxToInstr[i];
5320 // Ignore instructions that are never used within the loop.
5321 if (!Ends.count(I)) continue;
5323 // Remove all of the instructions that end at this location.
5324 InstrList &List = TransposeEnds[i];
5325 for (unsigned int j=0, e = List.size(); j < e; ++j)
5326 OpenIntervals.erase(List[j]);
5328 // Count the number of live interals.
5329 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5331 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5332 OpenIntervals.size() << '\n');
5334 // Add the current instruction to the list of open intervals.
5335 OpenIntervals.insert(I);
5338 unsigned Invariant = LoopInvariants.size();
5339 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5340 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5341 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5343 R.LoopInvariantRegs = Invariant;
5344 R.MaxLocalUsers = MaxUsage;
5348 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5352 for (Loop::block_iterator bb = TheLoop->block_begin(),
5353 be = TheLoop->block_end(); bb != be; ++bb) {
5354 unsigned BlockCost = 0;
5355 BasicBlock *BB = *bb;
5357 // For each instruction in the old loop.
5358 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5359 // Skip dbg intrinsics.
5360 if (isa<DbgInfoIntrinsic>(it))
5363 unsigned C = getInstructionCost(it, VF);
5365 // Check if we should override the cost.
5366 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5367 C = ForceTargetInstructionCost;
5370 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5371 VF << " For instruction: " << *it << '\n');
5374 // We assume that if-converted blocks have a 50% chance of being executed.
5375 // When the code is scalar then some of the blocks are avoided due to CF.
5376 // When the code is vectorized we execute all code paths.
5377 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5386 /// \brief Check whether the address computation for a non-consecutive memory
5387 /// access looks like an unlikely candidate for being merged into the indexing
5390 /// We look for a GEP which has one index that is an induction variable and all
5391 /// other indices are loop invariant. If the stride of this access is also
5392 /// within a small bound we decide that this address computation can likely be
5393 /// merged into the addressing mode.
5394 /// In all other cases, we identify the address computation as complex.
5395 static bool isLikelyComplexAddressComputation(Value *Ptr,
5396 LoopVectorizationLegality *Legal,
5397 ScalarEvolution *SE,
5398 const Loop *TheLoop) {
5399 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5403 // We are looking for a gep with all loop invariant indices except for one
5404 // which should be an induction variable.
5405 unsigned NumOperands = Gep->getNumOperands();
5406 for (unsigned i = 1; i < NumOperands; ++i) {
5407 Value *Opd = Gep->getOperand(i);
5408 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5409 !Legal->isInductionVariable(Opd))
5413 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5414 // can likely be merged into the address computation.
5415 unsigned MaxMergeDistance = 64;
5417 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5421 // Check the step is constant.
5422 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5423 // Calculate the pointer stride and check if it is consecutive.
5424 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5428 const APInt &APStepVal = C->getValue()->getValue();
5430 // Huge step value - give up.
5431 if (APStepVal.getBitWidth() > 64)
5434 int64_t StepVal = APStepVal.getSExtValue();
5436 return StepVal > MaxMergeDistance;
5439 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5440 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5446 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5447 // If we know that this instruction will remain uniform, check the cost of
5448 // the scalar version.
5449 if (Legal->isUniformAfterVectorization(I))
5452 Type *RetTy = I->getType();
5453 Type *VectorTy = ToVectorTy(RetTy, VF);
5455 // TODO: We need to estimate the cost of intrinsic calls.
5456 switch (I->getOpcode()) {
5457 case Instruction::GetElementPtr:
5458 // We mark this instruction as zero-cost because the cost of GEPs in
5459 // vectorized code depends on whether the corresponding memory instruction
5460 // is scalarized or not. Therefore, we handle GEPs with the memory
5461 // instruction cost.
5463 case Instruction::Br: {
5464 return TTI.getCFInstrCost(I->getOpcode());
5466 case Instruction::PHI:
5467 //TODO: IF-converted IFs become selects.
5469 case Instruction::Add:
5470 case Instruction::FAdd:
5471 case Instruction::Sub:
5472 case Instruction::FSub:
5473 case Instruction::Mul:
5474 case Instruction::FMul:
5475 case Instruction::UDiv:
5476 case Instruction::SDiv:
5477 case Instruction::FDiv:
5478 case Instruction::URem:
5479 case Instruction::SRem:
5480 case Instruction::FRem:
5481 case Instruction::Shl:
5482 case Instruction::LShr:
5483 case Instruction::AShr:
5484 case Instruction::And:
5485 case Instruction::Or:
5486 case Instruction::Xor: {
5487 // Since we will replace the stride by 1 the multiplication should go away.
5488 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5490 // Certain instructions can be cheaper to vectorize if they have a constant
5491 // second vector operand. One example of this are shifts on x86.
5492 TargetTransformInfo::OperandValueKind Op1VK =
5493 TargetTransformInfo::OK_AnyValue;
5494 TargetTransformInfo::OperandValueKind Op2VK =
5495 TargetTransformInfo::OK_AnyValue;
5496 Value *Op2 = I->getOperand(1);
5498 // Check for a splat of a constant or for a non uniform vector of constants.
5499 if (isa<ConstantInt>(Op2))
5500 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5501 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5502 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5503 if (cast<Constant>(Op2)->getSplatValue() != NULL)
5504 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5507 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5509 case Instruction::Select: {
5510 SelectInst *SI = cast<SelectInst>(I);
5511 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5512 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5513 Type *CondTy = SI->getCondition()->getType();
5515 CondTy = VectorType::get(CondTy, VF);
5517 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5519 case Instruction::ICmp:
5520 case Instruction::FCmp: {
5521 Type *ValTy = I->getOperand(0)->getType();
5522 VectorTy = ToVectorTy(ValTy, VF);
5523 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5525 case Instruction::Store:
5526 case Instruction::Load: {
5527 StoreInst *SI = dyn_cast<StoreInst>(I);
5528 LoadInst *LI = dyn_cast<LoadInst>(I);
5529 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5531 VectorTy = ToVectorTy(ValTy, VF);
5533 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5534 unsigned AS = SI ? SI->getPointerAddressSpace() :
5535 LI->getPointerAddressSpace();
5536 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5537 // We add the cost of address computation here instead of with the gep
5538 // instruction because only here we know whether the operation is
5541 return TTI.getAddressComputationCost(VectorTy) +
5542 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5544 // Scalarized loads/stores.
5545 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5546 bool Reverse = ConsecutiveStride < 0;
5547 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5548 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5549 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5550 bool IsComplexComputation =
5551 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5553 // The cost of extracting from the value vector and pointer vector.
5554 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5555 for (unsigned i = 0; i < VF; ++i) {
5556 // The cost of extracting the pointer operand.
5557 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5558 // In case of STORE, the cost of ExtractElement from the vector.
5559 // In case of LOAD, the cost of InsertElement into the returned
5561 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5562 Instruction::InsertElement,
5566 // The cost of the scalar loads/stores.
5567 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5568 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5573 // Wide load/stores.
5574 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5575 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5578 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5582 case Instruction::ZExt:
5583 case Instruction::SExt:
5584 case Instruction::FPToUI:
5585 case Instruction::FPToSI:
5586 case Instruction::FPExt:
5587 case Instruction::PtrToInt:
5588 case Instruction::IntToPtr:
5589 case Instruction::SIToFP:
5590 case Instruction::UIToFP:
5591 case Instruction::Trunc:
5592 case Instruction::FPTrunc:
5593 case Instruction::BitCast: {
5594 // We optimize the truncation of induction variable.
5595 // The cost of these is the same as the scalar operation.
5596 if (I->getOpcode() == Instruction::Trunc &&
5597 Legal->isInductionVariable(I->getOperand(0)))
5598 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5599 I->getOperand(0)->getType());
5601 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5602 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5604 case Instruction::Call: {
5605 CallInst *CI = cast<CallInst>(I);
5606 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5607 assert(ID && "Not an intrinsic call!");
5608 Type *RetTy = ToVectorTy(CI->getType(), VF);
5609 SmallVector<Type*, 4> Tys;
5610 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5611 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5612 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5615 // We are scalarizing the instruction. Return the cost of the scalar
5616 // instruction, plus the cost of insert and extract into vector
5617 // elements, times the vector width.
5620 if (!RetTy->isVoidTy() && VF != 1) {
5621 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5623 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5626 // The cost of inserting the results plus extracting each one of the
5628 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5631 // The cost of executing VF copies of the scalar instruction. This opcode
5632 // is unknown. Assume that it is the same as 'mul'.
5633 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5639 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5640 if (Scalar->isVoidTy() || VF == 1)
5642 return VectorType::get(Scalar, VF);
5645 char LoopVectorize::ID = 0;
5646 static const char lv_name[] = "Loop Vectorization";
5647 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5648 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5649 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5650 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5651 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5652 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5653 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5654 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5655 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5658 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5659 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5663 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5664 // Check for a store.
5665 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5666 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5668 // Check for a load.
5669 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5670 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5676 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5677 bool IfPredicateStore) {
5678 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5679 // Holds vector parameters or scalars, in case of uniform vals.
5680 SmallVector<VectorParts, 4> Params;
5682 setDebugLocFromInst(Builder, Instr);
5684 // Find all of the vectorized parameters.
5685 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5686 Value *SrcOp = Instr->getOperand(op);
5688 // If we are accessing the old induction variable, use the new one.
5689 if (SrcOp == OldInduction) {
5690 Params.push_back(getVectorValue(SrcOp));
5694 // Try using previously calculated values.
5695 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5697 // If the src is an instruction that appeared earlier in the basic block
5698 // then it should already be vectorized.
5699 if (SrcInst && OrigLoop->contains(SrcInst)) {
5700 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5701 // The parameter is a vector value from earlier.
5702 Params.push_back(WidenMap.get(SrcInst));
5704 // The parameter is a scalar from outside the loop. Maybe even a constant.
5705 VectorParts Scalars;
5706 Scalars.append(UF, SrcOp);
5707 Params.push_back(Scalars);
5711 assert(Params.size() == Instr->getNumOperands() &&
5712 "Invalid number of operands");
5714 // Does this instruction return a value ?
5715 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5717 Value *UndefVec = IsVoidRetTy ? 0 :
5718 UndefValue::get(Instr->getType());
5719 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5720 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5722 Instruction *InsertPt = Builder.GetInsertPoint();
5723 BasicBlock *IfBlock = Builder.GetInsertBlock();
5724 BasicBlock *CondBlock = 0;
5728 if (IfPredicateStore) {
5729 assert(Instr->getParent()->getSinglePredecessor() &&
5730 "Only support single predecessor blocks");
5731 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5732 Instr->getParent());
5733 VectorLp = LI->getLoopFor(IfBlock);
5734 assert(VectorLp && "Must have a loop for this block");
5737 // For each vector unroll 'part':
5738 for (unsigned Part = 0; Part < UF; ++Part) {
5739 // For each scalar that we create:
5741 // Start an "if (pred) a[i] = ..." block.
5743 if (IfPredicateStore) {
5744 if (Cond[Part]->getType()->isVectorTy())
5746 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5747 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5748 ConstantInt::get(Cond[Part]->getType(), 1));
5749 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5750 LoopVectorBody.push_back(CondBlock);
5751 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5752 // Update Builder with newly created basic block.
5753 Builder.SetInsertPoint(InsertPt);
5756 Instruction *Cloned = Instr->clone();
5758 Cloned->setName(Instr->getName() + ".cloned");
5759 // Replace the operands of the cloned instructions with extracted scalars.
5760 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5761 Value *Op = Params[op][Part];
5762 Cloned->setOperand(op, Op);
5765 // Place the cloned scalar in the new loop.
5766 Builder.Insert(Cloned);
5768 // If the original scalar returns a value we need to place it in a vector
5769 // so that future users will be able to use it.
5771 VecResults[Part] = Cloned;
5774 if (IfPredicateStore) {
5775 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5776 LoopVectorBody.push_back(NewIfBlock);
5777 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5778 Builder.SetInsertPoint(InsertPt);
5779 Instruction *OldBr = IfBlock->getTerminator();
5780 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5781 OldBr->eraseFromParent();
5782 IfBlock = NewIfBlock;
5787 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5788 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5789 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5791 return scalarizeInstruction(Instr, IfPredicateStore);
5794 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5798 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5802 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5804 // When unrolling and the VF is 1, we only need to add a simple scalar.
5805 Type *ITy = Val->getType();
5806 assert(!ITy->isVectorTy() && "Val must be a scalar");
5807 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5808 return Builder.CreateAdd(Val, C, "induction");