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/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Dominators.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/LLVMContext.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/IR/Type.h"
78 #include "llvm/IR/Value.h"
79 #include "llvm/IR/Verifier.h"
80 #include "llvm/Pass.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/Debug.h"
83 #include "llvm/Support/PatternMatch.h"
84 #include "llvm/Support/ValueHandle.h"
85 #include "llvm/Support/raw_ostream.h"
86 #include "llvm/Target/TargetLibraryInfo.h"
87 #include "llvm/Transforms/Scalar.h"
88 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
89 #include "llvm/Transforms/Utils/Local.h"
94 using namespace llvm::PatternMatch;
96 static cl::opt<unsigned>
97 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
98 cl::desc("Sets the SIMD width. Zero is autoselect."));
100 static cl::opt<unsigned>
101 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
102 cl::desc("Sets the vectorization unroll count. "
103 "Zero is autoselect."));
106 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
107 cl::desc("Enable if-conversion during vectorization."));
109 /// We don't vectorize loops with a known constant trip count below this number.
110 static cl::opt<unsigned>
111 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
113 cl::desc("Don't vectorize loops with a constant "
114 "trip count that is smaller than this "
117 /// This enables versioning on the strides of symbolically striding memory
118 /// accesses in code like the following.
119 /// for (i = 0; i < N; ++i)
120 /// A[i * Stride1] += B[i * Stride2] ...
122 /// Will be roughly translated to
123 /// if (Stride1 == 1 && Stride2 == 1) {
124 /// for (i = 0; i < N; i+=4)
128 static cl::opt<bool> EnableMemAccessVersioning(
129 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
130 cl::desc("Enable symblic stride memory access versioning"));
132 /// We don't unroll loops with a known constant trip count below this number.
133 static const unsigned TinyTripCountUnrollThreshold = 128;
135 /// When performing memory disambiguation checks at runtime do not make more
136 /// than this number of comparisons.
137 static const unsigned RuntimeMemoryCheckThreshold = 8;
139 /// Maximum simd width.
140 static const unsigned MaxVectorWidth = 64;
142 /// Maximum vectorization unroll count.
143 static const unsigned MaxUnrollFactor = 16;
145 /// The cost of a loop that is considered 'small' by the unroller.
146 static const unsigned SmallLoopCost = 20;
150 // Forward declarations.
151 class LoopVectorizationLegality;
152 class LoopVectorizationCostModel;
154 /// InnerLoopVectorizer vectorizes loops which contain only one basic
155 /// block to a specified vectorization factor (VF).
156 /// This class performs the widening of scalars into vectors, or multiple
157 /// scalars. This class also implements the following features:
158 /// * It inserts an epilogue loop for handling loops that don't have iteration
159 /// counts that are known to be a multiple of the vectorization factor.
160 /// * It handles the code generation for reduction variables.
161 /// * Scalarization (implementation using scalars) of un-vectorizable
163 /// InnerLoopVectorizer does not perform any vectorization-legality
164 /// checks, and relies on the caller to check for the different legality
165 /// aspects. The InnerLoopVectorizer relies on the
166 /// LoopVectorizationLegality class to provide information about the induction
167 /// and reduction variables that were found to a given vectorization factor.
168 class InnerLoopVectorizer {
170 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
171 DominatorTree *DT, DataLayout *DL,
172 const TargetLibraryInfo *TLI, unsigned VecWidth,
173 unsigned UnrollFactor)
174 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
175 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
176 OldInduction(0), WidenMap(UnrollFactor), Legal(0) {}
178 // Perform the actual loop widening (vectorization).
179 void vectorize(LoopVectorizationLegality *L) {
181 // Create a new empty loop. Unlink the old loop and connect the new one.
183 // Widen each instruction in the old loop to a new one in the new loop.
184 // Use the Legality module to find the induction and reduction variables.
186 // Register the new loop and update the analysis passes.
190 virtual ~InnerLoopVectorizer() {}
193 /// A small list of PHINodes.
194 typedef SmallVector<PHINode*, 4> PhiVector;
195 /// When we unroll loops we have multiple vector values for each scalar.
196 /// This data structure holds the unrolled and vectorized values that
197 /// originated from one scalar instruction.
198 typedef SmallVector<Value*, 2> VectorParts;
200 // When we if-convert we need create edge masks. We have to cache values so
201 // that we don't end up with exponential recursion/IR.
202 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
203 VectorParts> EdgeMaskCache;
205 /// \brief Add code that checks at runtime if the accessed arrays overlap.
207 /// Returns a pair of instructions where the first element is the first
208 /// instruction generated in possibly a sequence of instructions and the
209 /// second value is the final comparator value or NULL if no check is needed.
210 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
212 /// \brief Add checks for strides that where assumed to be 1.
214 /// Returns the last check instruction and the first check instruction in the
215 /// pair as (first, last).
216 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
218 /// Create an empty loop, based on the loop ranges of the old loop.
219 void createEmptyLoop();
220 /// Copy and widen the instructions from the old loop.
221 virtual void vectorizeLoop();
223 /// \brief The Loop exit block may have single value PHI nodes where the
224 /// incoming value is 'Undef'. While vectorizing we only handled real values
225 /// that were defined inside the loop. Here we fix the 'undef case'.
229 /// A helper function that computes the predicate of the block BB, assuming
230 /// that the header block of the loop is set to True. It returns the *entry*
231 /// mask for the block BB.
232 VectorParts createBlockInMask(BasicBlock *BB);
233 /// A helper function that computes the predicate of the edge between SRC
235 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
237 /// A helper function to vectorize a single BB within the innermost loop.
238 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
240 /// Vectorize a single PHINode in a block. This method handles the induction
241 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
242 /// arbitrary length vectors.
243 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
244 unsigned UF, unsigned VF, PhiVector *PV);
246 /// Insert the new loop to the loop hierarchy and pass manager
247 /// and update the analysis passes.
248 void updateAnalysis();
250 /// This instruction is un-vectorizable. Implement it as a sequence
252 virtual void scalarizeInstruction(Instruction *Instr);
254 /// Vectorize Load and Store instructions,
255 virtual void vectorizeMemoryInstruction(Instruction *Instr);
257 /// Create a broadcast instruction. This method generates a broadcast
258 /// instruction (shuffle) for loop invariant values and for the induction
259 /// value. If this is the induction variable then we extend it to N, N+1, ...
260 /// this is needed because each iteration in the loop corresponds to a SIMD
262 virtual Value *getBroadcastInstrs(Value *V);
264 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
265 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
266 /// The sequence starts at StartIndex.
267 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
269 /// When we go over instructions in the basic block we rely on previous
270 /// values within the current basic block or on loop invariant values.
271 /// When we widen (vectorize) values we place them in the map. If the values
272 /// are not within the map, they have to be loop invariant, so we simply
273 /// broadcast them into a vector.
274 VectorParts &getVectorValue(Value *V);
276 /// Generate a shuffle sequence that will reverse the vector Vec.
277 virtual Value *reverseVector(Value *Vec);
279 /// This is a helper class that holds the vectorizer state. It maps scalar
280 /// instructions to vector instructions. When the code is 'unrolled' then
281 /// then a single scalar value is mapped to multiple vector parts. The parts
282 /// are stored in the VectorPart type.
284 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
286 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
288 /// \return True if 'Key' is saved in the Value Map.
289 bool has(Value *Key) const { return MapStorage.count(Key); }
291 /// Initializes a new entry in the map. Sets all of the vector parts to the
292 /// save value in 'Val'.
293 /// \return A reference to a vector with splat values.
294 VectorParts &splat(Value *Key, Value *Val) {
295 VectorParts &Entry = MapStorage[Key];
296 Entry.assign(UF, Val);
300 ///\return A reference to the value that is stored at 'Key'.
301 VectorParts &get(Value *Key) {
302 VectorParts &Entry = MapStorage[Key];
305 assert(Entry.size() == UF);
310 /// The unroll factor. Each entry in the map stores this number of vector
314 /// Map storage. We use std::map and not DenseMap because insertions to a
315 /// dense map invalidates its iterators.
316 std::map<Value *, VectorParts> MapStorage;
319 /// The original loop.
321 /// Scev analysis to use.
329 /// Target Library Info.
330 const TargetLibraryInfo *TLI;
332 /// The vectorization SIMD factor to use. Each vector will have this many
337 /// The vectorization unroll factor to use. Each scalar is vectorized to this
338 /// many different vector instructions.
341 /// The builder that we use
344 // --- Vectorization state ---
346 /// The vector-loop preheader.
347 BasicBlock *LoopVectorPreHeader;
348 /// The scalar-loop preheader.
349 BasicBlock *LoopScalarPreHeader;
350 /// Middle Block between the vector and the scalar.
351 BasicBlock *LoopMiddleBlock;
352 ///The ExitBlock of the scalar loop.
353 BasicBlock *LoopExitBlock;
354 ///The vector loop body.
355 BasicBlock *LoopVectorBody;
356 ///The scalar loop body.
357 BasicBlock *LoopScalarBody;
358 /// A list of all bypass blocks. The first block is the entry of the loop.
359 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
361 /// The new Induction variable which was added to the new block.
363 /// The induction variable of the old basic block.
364 PHINode *OldInduction;
365 /// Holds the extended (to the widest induction type) start index.
367 /// Maps scalars to widened vectors.
369 EdgeMaskCache MaskCache;
371 LoopVectorizationLegality *Legal;
374 class InnerLoopUnroller : public InnerLoopVectorizer {
376 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
377 DominatorTree *DT, DataLayout *DL,
378 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
379 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
382 virtual void scalarizeInstruction(Instruction *Instr);
383 virtual void vectorizeMemoryInstruction(Instruction *Instr);
384 virtual Value *getBroadcastInstrs(Value *V);
385 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
386 virtual Value *reverseVector(Value *Vec);
389 /// \brief Look for a meaningful debug location on the instruction or it's
391 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
396 if (I->getDebugLoc() != Empty)
399 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
400 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
401 if (OpInst->getDebugLoc() != Empty)
408 /// \brief Set the debug location in the builder using the debug location in the
410 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
411 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
412 B.SetCurrentDebugLocation(Inst->getDebugLoc());
414 B.SetCurrentDebugLocation(DebugLoc());
417 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
418 /// to what vectorization factor.
419 /// This class does not look at the profitability of vectorization, only the
420 /// legality. This class has two main kinds of checks:
421 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
422 /// will change the order of memory accesses in a way that will change the
423 /// correctness of the program.
424 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
425 /// checks for a number of different conditions, such as the availability of a
426 /// single induction variable, that all types are supported and vectorize-able,
427 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
428 /// This class is also used by InnerLoopVectorizer for identifying
429 /// induction variable and the different reduction variables.
430 class LoopVectorizationLegality {
432 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
433 DominatorTree *DT, TargetLibraryInfo *TLI)
434 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
435 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
436 MaxSafeDepDistBytes(-1U) {}
438 /// This enum represents the kinds of reductions that we support.
440 RK_NoReduction, ///< Not a reduction.
441 RK_IntegerAdd, ///< Sum of integers.
442 RK_IntegerMult, ///< Product of integers.
443 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
444 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
445 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
446 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
447 RK_FloatAdd, ///< Sum of floats.
448 RK_FloatMult, ///< Product of floats.
449 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
452 /// This enum represents the kinds of inductions that we support.
454 IK_NoInduction, ///< Not an induction variable.
455 IK_IntInduction, ///< Integer induction variable. Step = 1.
456 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
457 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
458 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
461 // This enum represents the kind of minmax reduction.
462 enum MinMaxReductionKind {
472 /// This struct holds information about reduction variables.
473 struct ReductionDescriptor {
474 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
475 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
477 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
478 MinMaxReductionKind MK)
479 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
481 // The starting value of the reduction.
482 // It does not have to be zero!
483 TrackingVH<Value> StartValue;
484 // The instruction who's value is used outside the loop.
485 Instruction *LoopExitInstr;
486 // The kind of the reduction.
488 // If this a min/max reduction the kind of reduction.
489 MinMaxReductionKind MinMaxKind;
492 /// This POD struct holds information about a potential reduction operation.
493 struct ReductionInstDesc {
494 ReductionInstDesc(bool IsRedux, Instruction *I) :
495 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
497 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
498 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
500 // Is this instruction a reduction candidate.
502 // The last instruction in a min/max pattern (select of the select(icmp())
503 // pattern), or the current reduction instruction otherwise.
504 Instruction *PatternLastInst;
505 // If this is a min/max pattern the comparison predicate.
506 MinMaxReductionKind MinMaxKind;
509 /// This struct holds information about the memory runtime legality
510 /// check that a group of pointers do not overlap.
511 struct RuntimePointerCheck {
512 RuntimePointerCheck() : Need(false) {}
514 /// Reset the state of the pointer runtime information.
521 DependencySetId.clear();
524 /// Insert a pointer and calculate the start and end SCEVs.
525 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
526 unsigned DepSetId, ValueToValueMap &Strides);
528 /// This flag indicates if we need to add the runtime check.
530 /// Holds the pointers that we need to check.
531 SmallVector<TrackingVH<Value>, 2> Pointers;
532 /// Holds the pointer value at the beginning of the loop.
533 SmallVector<const SCEV*, 2> Starts;
534 /// Holds the pointer value at the end of the loop.
535 SmallVector<const SCEV*, 2> Ends;
536 /// Holds the information if this pointer is used for writing to memory.
537 SmallVector<bool, 2> IsWritePtr;
538 /// Holds the id of the set of pointers that could be dependent because of a
539 /// shared underlying object.
540 SmallVector<unsigned, 2> DependencySetId;
543 /// A struct for saving information about induction variables.
544 struct InductionInfo {
545 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
546 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
548 TrackingVH<Value> StartValue;
553 /// ReductionList contains the reduction descriptors for all
554 /// of the reductions that were found in the loop.
555 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
557 /// InductionList saves induction variables and maps them to the
558 /// induction descriptor.
559 typedef MapVector<PHINode*, InductionInfo> InductionList;
561 /// Returns true if it is legal to vectorize this loop.
562 /// This does not mean that it is profitable to vectorize this
563 /// loop, only that it is legal to do so.
566 /// Returns the Induction variable.
567 PHINode *getInduction() { return Induction; }
569 /// Returns the reduction variables found in the loop.
570 ReductionList *getReductionVars() { return &Reductions; }
572 /// Returns the induction variables found in the loop.
573 InductionList *getInductionVars() { return &Inductions; }
575 /// Returns the widest induction type.
576 Type *getWidestInductionType() { return WidestIndTy; }
578 /// Returns True if V is an induction variable in this loop.
579 bool isInductionVariable(const Value *V);
581 /// Return true if the block BB needs to be predicated in order for the loop
582 /// to be vectorized.
583 bool blockNeedsPredication(BasicBlock *BB);
585 /// Check if this pointer is consecutive when vectorizing. This happens
586 /// when the last index of the GEP is the induction variable, or that the
587 /// pointer itself is an induction variable.
588 /// This check allows us to vectorize A[idx] into a wide load/store.
590 /// 0 - Stride is unknown or non-consecutive.
591 /// 1 - Address is consecutive.
592 /// -1 - Address is consecutive, and decreasing.
593 int isConsecutivePtr(Value *Ptr);
595 /// Returns true if the value V is uniform within the loop.
596 bool isUniform(Value *V);
598 /// Returns true if this instruction will remain scalar after vectorization.
599 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
601 /// Returns the information that we collected about runtime memory check.
602 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
604 /// This function returns the identity element (or neutral element) for
606 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
608 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
610 bool hasStride(Value *V) { return StrideSet.count(V); }
611 bool mustCheckStrides() { return !StrideSet.empty(); }
612 SmallPtrSet<Value *, 8>::iterator strides_begin() {
613 return StrideSet.begin();
615 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
618 /// Check if a single basic block loop is vectorizable.
619 /// At this point we know that this is a loop with a constant trip count
620 /// and we only need to check individual instructions.
621 bool canVectorizeInstrs();
623 /// When we vectorize loops we may change the order in which
624 /// we read and write from memory. This method checks if it is
625 /// legal to vectorize the code, considering only memory constrains.
626 /// Returns true if the loop is vectorizable
627 bool canVectorizeMemory();
629 /// Return true if we can vectorize this loop using the IF-conversion
631 bool canVectorizeWithIfConvert();
633 /// Collect the variables that need to stay uniform after vectorization.
634 void collectLoopUniforms();
636 /// Return true if all of the instructions in the block can be speculatively
637 /// executed. \p SafePtrs is a list of addresses that are known to be legal
638 /// and we know that we can read from them without segfault.
639 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
641 /// Returns True, if 'Phi' is the kind of reduction variable for type
642 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
643 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
644 /// Returns a struct describing if the instruction 'I' can be a reduction
645 /// variable of type 'Kind'. If the reduction is a min/max pattern of
646 /// select(icmp()) this function advances the instruction pointer 'I' from the
647 /// compare instruction to the select instruction and stores this pointer in
648 /// 'PatternLastInst' member of the returned struct.
649 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
650 ReductionInstDesc &Desc);
651 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
652 /// pattern corresponding to a min(X, Y) or max(X, Y).
653 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
654 ReductionInstDesc &Prev);
655 /// Returns the induction kind of Phi. This function may return NoInduction
656 /// if the PHI is not an induction variable.
657 InductionKind isInductionVariable(PHINode *Phi);
659 /// \brief Collect memory access with loop invariant strides.
661 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
663 void collectStridedAcccess(Value *LoadOrStoreInst);
665 /// The loop that we evaluate.
669 /// DataLayout analysis.
673 /// Target Library Info.
674 TargetLibraryInfo *TLI;
676 // --- vectorization state --- //
678 /// Holds the integer induction variable. This is the counter of the
681 /// Holds the reduction variables.
682 ReductionList Reductions;
683 /// Holds all of the induction variables that we found in the loop.
684 /// Notice that inductions don't need to start at zero and that induction
685 /// variables can be pointers.
686 InductionList Inductions;
687 /// Holds the widest induction type encountered.
690 /// Allowed outside users. This holds the reduction
691 /// vars which can be accessed from outside the loop.
692 SmallPtrSet<Value*, 4> AllowedExit;
693 /// This set holds the variables which are known to be uniform after
695 SmallPtrSet<Instruction*, 4> Uniforms;
696 /// We need to check that all of the pointers in this list are disjoint
698 RuntimePointerCheck PtrRtCheck;
699 /// Can we assume the absence of NaNs.
700 bool HasFunNoNaNAttr;
702 unsigned MaxSafeDepDistBytes;
704 ValueToValueMap Strides;
705 SmallPtrSet<Value *, 8> StrideSet;
708 /// LoopVectorizationCostModel - estimates the expected speedups due to
710 /// In many cases vectorization is not profitable. This can happen because of
711 /// a number of reasons. In this class we mainly attempt to predict the
712 /// expected speedup/slowdowns due to the supported instruction set. We use the
713 /// TargetTransformInfo to query the different backends for the cost of
714 /// different operations.
715 class LoopVectorizationCostModel {
717 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
718 LoopVectorizationLegality *Legal,
719 const TargetTransformInfo &TTI,
720 DataLayout *DL, const TargetLibraryInfo *TLI)
721 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
723 /// Information about vectorization costs
724 struct VectorizationFactor {
725 unsigned Width; // Vector width with best cost
726 unsigned Cost; // Cost of the loop with that width
728 /// \return The most profitable vectorization factor and the cost of that VF.
729 /// This method checks every power of two up to VF. If UserVF is not ZERO
730 /// then this vectorization factor will be selected if vectorization is
732 VectorizationFactor selectVectorizationFactor(bool OptForSize,
735 /// \return The size (in bits) of the widest type in the code that
736 /// needs to be vectorized. We ignore values that remain scalar such as
737 /// 64 bit loop indices.
738 unsigned getWidestType();
740 /// \return The most profitable unroll factor.
741 /// If UserUF is non-zero then this method finds the best unroll-factor
742 /// based on register pressure and other parameters.
743 /// VF and LoopCost are the selected vectorization factor and the cost of the
745 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
748 /// \brief A struct that represents some properties of the register usage
750 struct RegisterUsage {
751 /// Holds the number of loop invariant values that are used in the loop.
752 unsigned LoopInvariantRegs;
753 /// Holds the maximum number of concurrent live intervals in the loop.
754 unsigned MaxLocalUsers;
755 /// Holds the number of instructions in the loop.
756 unsigned NumInstructions;
759 /// \return information about the register usage of the loop.
760 RegisterUsage calculateRegisterUsage();
763 /// Returns the expected execution cost. The unit of the cost does
764 /// not matter because we use the 'cost' units to compare different
765 /// vector widths. The cost that is returned is *not* normalized by
766 /// the factor width.
767 unsigned expectedCost(unsigned VF);
769 /// Returns the execution time cost of an instruction for a given vector
770 /// width. Vector width of one means scalar.
771 unsigned getInstructionCost(Instruction *I, unsigned VF);
773 /// A helper function for converting Scalar types to vector types.
774 /// If the incoming type is void, we return void. If the VF is 1, we return
776 static Type* ToVectorTy(Type *Scalar, unsigned VF);
778 /// Returns whether the instruction is a load or store and will be a emitted
779 /// as a vector operation.
780 bool isConsecutiveLoadOrStore(Instruction *I);
782 /// The loop that we evaluate.
786 /// Loop Info analysis.
788 /// Vectorization legality.
789 LoopVectorizationLegality *Legal;
790 /// Vector target information.
791 const TargetTransformInfo &TTI;
792 /// Target data layout information.
794 /// Target Library Info.
795 const TargetLibraryInfo *TLI;
798 /// Utility class for getting and setting loop vectorizer hints in the form
799 /// of loop metadata.
800 struct LoopVectorizeHints {
801 /// Vectorization width.
803 /// Vectorization unroll factor.
805 /// Vectorization forced (-1 not selected, 0 force disabled, 1 force enabled)
808 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
809 : Width(VectorizationFactor)
810 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
812 , LoopID(L->getLoopID()) {
814 // The command line options override any loop metadata except for when
815 // width == 1 which is used to indicate the loop is already vectorized.
816 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
817 Width = VectorizationFactor;
818 if (VectorizationUnroll.getNumOccurrences() > 0)
819 Unroll = VectorizationUnroll;
821 DEBUG(if (DisableUnrolling && Unroll == 1)
822 dbgs() << "LV: Unrolling disabled by the pass manager\n");
825 /// Return the loop vectorizer metadata prefix.
826 static StringRef Prefix() { return "llvm.vectorizer."; }
828 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
829 SmallVector<Value*, 2> Vals;
830 Vals.push_back(MDString::get(Context, Name));
831 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
832 return MDNode::get(Context, Vals);
835 /// Mark the loop L as already vectorized by setting the width to 1.
836 void setAlreadyVectorized(Loop *L) {
837 LLVMContext &Context = L->getHeader()->getContext();
841 // Create a new loop id with one more operand for the already_vectorized
842 // hint. If the loop already has a loop id then copy the existing operands.
843 SmallVector<Value*, 4> Vals(1);
845 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
846 Vals.push_back(LoopID->getOperand(i));
848 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
849 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
851 MDNode *NewLoopID = MDNode::get(Context, Vals);
852 // Set operand 0 to refer to the loop id itself.
853 NewLoopID->replaceOperandWith(0, NewLoopID);
855 L->setLoopID(NewLoopID);
857 LoopID->replaceAllUsesWith(NewLoopID);
865 /// Find hints specified in the loop metadata.
866 void getHints(const Loop *L) {
870 // First operand should refer to the loop id itself.
871 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
872 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
874 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
875 const MDString *S = 0;
876 SmallVector<Value*, 4> Args;
878 // The expected hint is either a MDString or a MDNode with the first
879 // operand a MDString.
880 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
881 if (!MD || MD->getNumOperands() == 0)
883 S = dyn_cast<MDString>(MD->getOperand(0));
884 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
885 Args.push_back(MD->getOperand(i));
887 S = dyn_cast<MDString>(LoopID->getOperand(i));
888 assert(Args.size() == 0 && "too many arguments for MDString");
894 // Check if the hint starts with the vectorizer prefix.
895 StringRef Hint = S->getString();
896 if (!Hint.startswith(Prefix()))
898 // Remove the prefix.
899 Hint = Hint.substr(Prefix().size(), StringRef::npos);
901 if (Args.size() == 1)
902 getHint(Hint, Args[0]);
906 // Check string hint with one operand.
907 void getHint(StringRef Hint, Value *Arg) {
908 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
910 unsigned Val = C->getZExtValue();
912 if (Hint == "width") {
913 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
916 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
917 } else if (Hint == "unroll") {
918 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
921 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
922 } else if (Hint == "enable") {
923 if (C->getBitWidth() == 1)
926 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
928 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
933 /// The LoopVectorize Pass.
934 struct LoopVectorize : public LoopPass {
935 /// Pass identification, replacement for typeid
938 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
940 DisableUnrolling(NoUnrolling),
941 AlwaysVectorize(AlwaysVectorize) {
942 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
948 TargetTransformInfo *TTI;
950 TargetLibraryInfo *TLI;
951 bool DisableUnrolling;
952 bool AlwaysVectorize;
954 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
955 // We only vectorize innermost loops.
959 SE = &getAnalysis<ScalarEvolution>();
960 DL = getAnalysisIfAvailable<DataLayout>();
961 LI = &getAnalysis<LoopInfo>();
962 TTI = &getAnalysis<TargetTransformInfo>();
963 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
964 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
966 // If the target claims to have no vector registers don't attempt
968 if (!TTI->getNumberOfRegisters(true))
972 DEBUG(dbgs() << "LV: Not vectorizing: Missing data layout\n");
976 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
977 L->getHeader()->getParent()->getName() << "\"\n");
979 LoopVectorizeHints Hints(L, DisableUnrolling);
981 if (Hints.Force == 0) {
982 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
986 if (!AlwaysVectorize && Hints.Force != 1) {
987 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
991 if (Hints.Width == 1 && Hints.Unroll == 1) {
992 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
996 // Check if it is legal to vectorize the loop.
997 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
998 if (!LVL.canVectorize()) {
999 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1003 // Use the cost model.
1004 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1006 // Check the function attributes to find out if this function should be
1007 // optimized for size.
1008 Function *F = L->getHeader()->getParent();
1009 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
1010 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
1011 unsigned FnIndex = AttributeSet::FunctionIndex;
1012 bool OptForSize = Hints.Force != 1 &&
1013 F->getAttributes().hasAttribute(FnIndex, SzAttr);
1014 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
1017 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1018 "attribute is used.\n");
1022 // Select the optimal vectorization factor.
1023 LoopVectorizationCostModel::VectorizationFactor VF;
1024 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
1025 // Select the unroll factor.
1026 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1029 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
1030 F->getParent()->getModuleIdentifier() << '\n');
1031 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1033 if (VF.Width == 1) {
1034 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1037 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1038 // We decided not to vectorize, but we may want to unroll.
1039 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1040 Unroller.vectorize(&LVL);
1042 // If we decided that it is *legal* to vectorize the loop then do it.
1043 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1047 // Mark the loop as already vectorized to avoid vectorizing again.
1048 Hints.setAlreadyVectorized(L);
1050 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1054 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1055 LoopPass::getAnalysisUsage(AU);
1056 AU.addRequiredID(LoopSimplifyID);
1057 AU.addRequiredID(LCSSAID);
1058 AU.addRequired<DominatorTreeWrapperPass>();
1059 AU.addRequired<LoopInfo>();
1060 AU.addRequired<ScalarEvolution>();
1061 AU.addRequired<TargetTransformInfo>();
1062 AU.addPreserved<LoopInfo>();
1063 AU.addPreserved<DominatorTreeWrapperPass>();
1068 } // end anonymous namespace
1070 //===----------------------------------------------------------------------===//
1071 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1072 // LoopVectorizationCostModel.
1073 //===----------------------------------------------------------------------===//
1075 static Value *stripIntegerCast(Value *V) {
1076 if (CastInst *CI = dyn_cast<CastInst>(V))
1077 if (CI->getOperand(0)->getType()->isIntegerTy())
1078 return CI->getOperand(0);
1082 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1084 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1086 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1087 ValueToValueMap &PtrToStride,
1088 Value *Ptr, Value *OrigPtr = 0) {
1090 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1092 // If there is an entry in the map return the SCEV of the pointer with the
1093 // symbolic stride replaced by one.
1094 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1095 if (SI != PtrToStride.end()) {
1096 Value *StrideVal = SI->second;
1099 StrideVal = stripIntegerCast(StrideVal);
1101 // Replace symbolic stride by one.
1102 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1103 ValueToValueMap RewriteMap;
1104 RewriteMap[StrideVal] = One;
1107 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1108 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1113 // Otherwise, just return the SCEV of the original pointer.
1114 return SE->getSCEV(Ptr);
1117 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1118 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1119 ValueToValueMap &Strides) {
1120 // Get the stride replaced scev.
1121 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1122 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1123 assert(AR && "Invalid addrec expression");
1124 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1125 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1126 Pointers.push_back(Ptr);
1127 Starts.push_back(AR->getStart());
1128 Ends.push_back(ScEnd);
1129 IsWritePtr.push_back(WritePtr);
1130 DependencySetId.push_back(DepSetId);
1133 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1134 // We need to place the broadcast of invariant variables outside the loop.
1135 Instruction *Instr = dyn_cast<Instruction>(V);
1136 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1137 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1139 // Place the code for broadcasting invariant variables in the new preheader.
1140 IRBuilder<>::InsertPointGuard Guard(Builder);
1142 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1144 // Broadcast the scalar into all locations in the vector.
1145 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1150 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1152 assert(Val->getType()->isVectorTy() && "Must be a vector");
1153 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1154 "Elem must be an integer");
1155 // Create the types.
1156 Type *ITy = Val->getType()->getScalarType();
1157 VectorType *Ty = cast<VectorType>(Val->getType());
1158 int VLen = Ty->getNumElements();
1159 SmallVector<Constant*, 8> Indices;
1161 // Create a vector of consecutive numbers from zero to VF.
1162 for (int i = 0; i < VLen; ++i) {
1163 int64_t Idx = Negate ? (-i) : i;
1164 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1167 // Add the consecutive indices to the vector value.
1168 Constant *Cv = ConstantVector::get(Indices);
1169 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1170 return Builder.CreateAdd(Val, Cv, "induction");
1173 /// \brief Find the operand of the GEP that should be checked for consecutive
1174 /// stores. This ignores trailing indices that have no effect on the final
1176 static unsigned getGEPInductionOperand(DataLayout *DL,
1177 const GetElementPtrInst *Gep) {
1178 unsigned LastOperand = Gep->getNumOperands() - 1;
1179 unsigned GEPAllocSize = DL->getTypeAllocSize(
1180 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1182 // Walk backwards and try to peel off zeros.
1183 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1184 // Find the type we're currently indexing into.
1185 gep_type_iterator GEPTI = gep_type_begin(Gep);
1186 std::advance(GEPTI, LastOperand - 1);
1188 // If it's a type with the same allocation size as the result of the GEP we
1189 // can peel off the zero index.
1190 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1198 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1199 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1200 // Make sure that the pointer does not point to structs.
1201 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1204 // If this value is a pointer induction variable we know it is consecutive.
1205 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1206 if (Phi && Inductions.count(Phi)) {
1207 InductionInfo II = Inductions[Phi];
1208 if (IK_PtrInduction == II.IK)
1210 else if (IK_ReversePtrInduction == II.IK)
1214 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1218 unsigned NumOperands = Gep->getNumOperands();
1219 Value *GpPtr = Gep->getPointerOperand();
1220 // If this GEP value is a consecutive pointer induction variable and all of
1221 // the indices are constant then we know it is consecutive. We can
1222 Phi = dyn_cast<PHINode>(GpPtr);
1223 if (Phi && Inductions.count(Phi)) {
1225 // Make sure that the pointer does not point to structs.
1226 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1227 if (GepPtrType->getElementType()->isAggregateType())
1230 // Make sure that all of the index operands are loop invariant.
1231 for (unsigned i = 1; i < NumOperands; ++i)
1232 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1235 InductionInfo II = Inductions[Phi];
1236 if (IK_PtrInduction == II.IK)
1238 else if (IK_ReversePtrInduction == II.IK)
1242 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1244 // Check that all of the gep indices are uniform except for our induction
1246 for (unsigned i = 0; i != NumOperands; ++i)
1247 if (i != InductionOperand &&
1248 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1251 // We can emit wide load/stores only if the last non-zero index is the
1252 // induction variable.
1253 const SCEV *Last = 0;
1254 if (!Strides.count(Gep))
1255 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1257 // Because of the multiplication by a stride we can have a s/zext cast.
1258 // We are going to replace this stride by 1 so the cast is safe to ignore.
1260 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1261 // %0 = trunc i64 %indvars.iv to i32
1262 // %mul = mul i32 %0, %Stride1
1263 // %idxprom = zext i32 %mul to i64 << Safe cast.
1264 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1266 Last = replaceSymbolicStrideSCEV(SE, Strides,
1267 Gep->getOperand(InductionOperand), Gep);
1268 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1270 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1274 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1275 const SCEV *Step = AR->getStepRecurrence(*SE);
1277 // The memory is consecutive because the last index is consecutive
1278 // and all other indices are loop invariant.
1281 if (Step->isAllOnesValue())
1288 bool LoopVectorizationLegality::isUniform(Value *V) {
1289 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1292 InnerLoopVectorizer::VectorParts&
1293 InnerLoopVectorizer::getVectorValue(Value *V) {
1294 assert(V != Induction && "The new induction variable should not be used.");
1295 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1297 // If we have a stride that is replaced by one, do it here.
1298 if (Legal->hasStride(V))
1299 V = ConstantInt::get(V->getType(), 1);
1301 // If we have this scalar in the map, return it.
1302 if (WidenMap.has(V))
1303 return WidenMap.get(V);
1305 // If this scalar is unknown, assume that it is a constant or that it is
1306 // loop invariant. Broadcast V and save the value for future uses.
1307 Value *B = getBroadcastInstrs(V);
1308 return WidenMap.splat(V, B);
1311 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1312 assert(Vec->getType()->isVectorTy() && "Invalid type");
1313 SmallVector<Constant*, 8> ShuffleMask;
1314 for (unsigned i = 0; i < VF; ++i)
1315 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1317 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1318 ConstantVector::get(ShuffleMask),
1322 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1323 // Attempt to issue a wide load.
1324 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1325 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1327 assert((LI || SI) && "Invalid Load/Store instruction");
1329 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1330 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1331 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1332 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1333 // An alignment of 0 means target abi alignment. We need to use the scalar's
1334 // target abi alignment in such a case.
1336 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1337 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1338 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1339 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1341 if (ScalarAllocatedSize != VectorElementSize)
1342 return scalarizeInstruction(Instr);
1344 // If the pointer is loop invariant or if it is non-consecutive,
1345 // scalarize the load.
1346 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1347 bool Reverse = ConsecutiveStride < 0;
1348 bool UniformLoad = LI && Legal->isUniform(Ptr);
1349 if (!ConsecutiveStride || UniformLoad)
1350 return scalarizeInstruction(Instr);
1352 Constant *Zero = Builder.getInt32(0);
1353 VectorParts &Entry = WidenMap.get(Instr);
1355 // Handle consecutive loads/stores.
1356 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1357 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1358 setDebugLocFromInst(Builder, Gep);
1359 Value *PtrOperand = Gep->getPointerOperand();
1360 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1361 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1363 // Create the new GEP with the new induction variable.
1364 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1365 Gep2->setOperand(0, FirstBasePtr);
1366 Gep2->setName("gep.indvar.base");
1367 Ptr = Builder.Insert(Gep2);
1369 setDebugLocFromInst(Builder, Gep);
1370 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1371 OrigLoop) && "Base ptr must be invariant");
1373 // The last index does not have to be the induction. It can be
1374 // consecutive and be a function of the index. For example A[I+1];
1375 unsigned NumOperands = Gep->getNumOperands();
1376 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1377 // Create the new GEP with the new induction variable.
1378 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1380 for (unsigned i = 0; i < NumOperands; ++i) {
1381 Value *GepOperand = Gep->getOperand(i);
1382 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1384 // Update last index or loop invariant instruction anchored in loop.
1385 if (i == InductionOperand ||
1386 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1387 assert((i == InductionOperand ||
1388 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1389 "Must be last index or loop invariant");
1391 VectorParts &GEPParts = getVectorValue(GepOperand);
1392 Value *Index = GEPParts[0];
1393 Index = Builder.CreateExtractElement(Index, Zero);
1394 Gep2->setOperand(i, Index);
1395 Gep2->setName("gep.indvar.idx");
1398 Ptr = Builder.Insert(Gep2);
1400 // Use the induction element ptr.
1401 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1402 setDebugLocFromInst(Builder, Ptr);
1403 VectorParts &PtrVal = getVectorValue(Ptr);
1404 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1409 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1410 "We do not allow storing to uniform addresses");
1411 setDebugLocFromInst(Builder, SI);
1412 // We don't want to update the value in the map as it might be used in
1413 // another expression. So don't use a reference type for "StoredVal".
1414 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1416 for (unsigned Part = 0; Part < UF; ++Part) {
1417 // Calculate the pointer for the specific unroll-part.
1418 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1421 // If we store to reverse consecutive memory locations then we need
1422 // to reverse the order of elements in the stored value.
1423 StoredVal[Part] = reverseVector(StoredVal[Part]);
1424 // If the address is consecutive but reversed, then the
1425 // wide store needs to start at the last vector element.
1426 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1427 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1430 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1431 DataTy->getPointerTo(AddressSpace));
1432 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1438 assert(LI && "Must have a load instruction");
1439 setDebugLocFromInst(Builder, LI);
1440 for (unsigned Part = 0; Part < UF; ++Part) {
1441 // Calculate the pointer for the specific unroll-part.
1442 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1445 // If the address is consecutive but reversed, then the
1446 // wide store needs to start at the last vector element.
1447 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1448 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1451 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1452 DataTy->getPointerTo(AddressSpace));
1453 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1454 cast<LoadInst>(LI)->setAlignment(Alignment);
1455 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1459 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1460 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1461 // Holds vector parameters or scalars, in case of uniform vals.
1462 SmallVector<VectorParts, 4> Params;
1464 setDebugLocFromInst(Builder, Instr);
1466 // Find all of the vectorized parameters.
1467 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1468 Value *SrcOp = Instr->getOperand(op);
1470 // If we are accessing the old induction variable, use the new one.
1471 if (SrcOp == OldInduction) {
1472 Params.push_back(getVectorValue(SrcOp));
1476 // Try using previously calculated values.
1477 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1479 // If the src is an instruction that appeared earlier in the basic block
1480 // then it should already be vectorized.
1481 if (SrcInst && OrigLoop->contains(SrcInst)) {
1482 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1483 // The parameter is a vector value from earlier.
1484 Params.push_back(WidenMap.get(SrcInst));
1486 // The parameter is a scalar from outside the loop. Maybe even a constant.
1487 VectorParts Scalars;
1488 Scalars.append(UF, SrcOp);
1489 Params.push_back(Scalars);
1493 assert(Params.size() == Instr->getNumOperands() &&
1494 "Invalid number of operands");
1496 // Does this instruction return a value ?
1497 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1499 Value *UndefVec = IsVoidRetTy ? 0 :
1500 UndefValue::get(VectorType::get(Instr->getType(), VF));
1501 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1502 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1504 // For each vector unroll 'part':
1505 for (unsigned Part = 0; Part < UF; ++Part) {
1506 // For each scalar that we create:
1507 for (unsigned Width = 0; Width < VF; ++Width) {
1508 Instruction *Cloned = Instr->clone();
1510 Cloned->setName(Instr->getName() + ".cloned");
1511 // Replace the operands of the cloned instructions with extracted scalars.
1512 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1513 Value *Op = Params[op][Part];
1514 // Param is a vector. Need to extract the right lane.
1515 if (Op->getType()->isVectorTy())
1516 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1517 Cloned->setOperand(op, Op);
1520 // Place the cloned scalar in the new loop.
1521 Builder.Insert(Cloned);
1523 // If the original scalar returns a value we need to place it in a vector
1524 // so that future users will be able to use it.
1526 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1527 Builder.getInt32(Width));
1532 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1536 if (Instruction *I = dyn_cast<Instruction>(V))
1537 return I->getParent() == Loc->getParent() ? I : 0;
1541 std::pair<Instruction *, Instruction *>
1542 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1543 Instruction *tnullptr = 0;
1544 if (!Legal->mustCheckStrides())
1545 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1547 IRBuilder<> ChkBuilder(Loc);
1551 Instruction *FirstInst = 0;
1552 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1553 SE = Legal->strides_end();
1555 Value *Ptr = stripIntegerCast(*SI);
1556 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1558 // Store the first instruction we create.
1559 FirstInst = getFirstInst(FirstInst, C, Loc);
1561 Check = ChkBuilder.CreateOr(Check, C);
1566 // We have to do this trickery because the IRBuilder might fold the check to a
1567 // constant expression in which case there is no Instruction anchored in a
1569 LLVMContext &Ctx = Loc->getContext();
1570 Instruction *TheCheck =
1571 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1572 ChkBuilder.Insert(TheCheck, "stride.not.one");
1573 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1575 return std::make_pair(FirstInst, TheCheck);
1578 std::pair<Instruction *, Instruction *>
1579 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1580 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1581 Legal->getRuntimePointerCheck();
1583 Instruction *tnullptr = 0;
1584 if (!PtrRtCheck->Need)
1585 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1587 unsigned NumPointers = PtrRtCheck->Pointers.size();
1588 SmallVector<TrackingVH<Value> , 2> Starts;
1589 SmallVector<TrackingVH<Value> , 2> Ends;
1591 LLVMContext &Ctx = Loc->getContext();
1592 SCEVExpander Exp(*SE, "induction");
1593 Instruction *FirstInst = 0;
1595 for (unsigned i = 0; i < NumPointers; ++i) {
1596 Value *Ptr = PtrRtCheck->Pointers[i];
1597 const SCEV *Sc = SE->getSCEV(Ptr);
1599 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1600 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1602 Starts.push_back(Ptr);
1603 Ends.push_back(Ptr);
1605 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1606 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1608 // Use this type for pointer arithmetic.
1609 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1611 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1612 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1613 Starts.push_back(Start);
1614 Ends.push_back(End);
1618 IRBuilder<> ChkBuilder(Loc);
1619 // Our instructions might fold to a constant.
1620 Value *MemoryRuntimeCheck = 0;
1621 for (unsigned i = 0; i < NumPointers; ++i) {
1622 for (unsigned j = i+1; j < NumPointers; ++j) {
1623 // No need to check if two readonly pointers intersect.
1624 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1627 // Only need to check pointers between two different dependency sets.
1628 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1631 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1632 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1634 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1635 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1636 "Trying to bounds check pointers with different address spaces");
1638 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1639 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1641 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1642 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1643 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1644 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1646 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1647 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1648 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1649 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1650 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1651 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1652 if (MemoryRuntimeCheck) {
1653 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1655 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1657 MemoryRuntimeCheck = IsConflict;
1661 // We have to do this trickery because the IRBuilder might fold the check to a
1662 // constant expression in which case there is no Instruction anchored in a
1664 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1665 ConstantInt::getTrue(Ctx));
1666 ChkBuilder.Insert(Check, "memcheck.conflict");
1667 FirstInst = getFirstInst(FirstInst, Check, Loc);
1668 return std::make_pair(FirstInst, Check);
1671 void InnerLoopVectorizer::createEmptyLoop() {
1673 In this function we generate a new loop. The new loop will contain
1674 the vectorized instructions while the old loop will continue to run the
1677 [ ] <-- vector loop bypass (may consist of multiple blocks).
1680 | [ ] <-- vector pre header.
1684 | [ ]_| <-- vector loop.
1687 >[ ] <--- middle-block.
1690 | [ ] <--- new preheader.
1694 | [ ]_| <-- old scalar loop to handle remainder.
1697 >[ ] <-- exit block.
1701 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1702 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1703 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1704 assert(ExitBlock && "Must have an exit block");
1706 // Some loops have a single integer induction variable, while other loops
1707 // don't. One example is c++ iterators that often have multiple pointer
1708 // induction variables. In the code below we also support a case where we
1709 // don't have a single induction variable.
1710 OldInduction = Legal->getInduction();
1711 Type *IdxTy = Legal->getWidestInductionType();
1713 // Find the loop boundaries.
1714 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1715 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1717 // The exit count might have the type of i64 while the phi is i32. This can
1718 // happen if we have an induction variable that is sign extended before the
1719 // compare. The only way that we get a backedge taken count is that the
1720 // induction variable was signed and as such will not overflow. In such a case
1721 // truncation is legal.
1722 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1723 IdxTy->getPrimitiveSizeInBits())
1724 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1726 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1727 // Get the total trip count from the count by adding 1.
1728 ExitCount = SE->getAddExpr(ExitCount,
1729 SE->getConstant(ExitCount->getType(), 1));
1731 // Expand the trip count and place the new instructions in the preheader.
1732 // Notice that the pre-header does not change, only the loop body.
1733 SCEVExpander Exp(*SE, "induction");
1735 // Count holds the overall loop count (N).
1736 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1737 BypassBlock->getTerminator());
1739 // The loop index does not have to start at Zero. Find the original start
1740 // value from the induction PHI node. If we don't have an induction variable
1741 // then we know that it starts at zero.
1742 Builder.SetInsertPoint(BypassBlock->getTerminator());
1743 Value *StartIdx = ExtendedIdx = OldInduction ?
1744 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1746 ConstantInt::get(IdxTy, 0);
1748 assert(BypassBlock && "Invalid loop structure");
1749 LoopBypassBlocks.push_back(BypassBlock);
1751 // Split the single block loop into the two loop structure described above.
1752 BasicBlock *VectorPH =
1753 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1754 BasicBlock *VecBody =
1755 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1756 BasicBlock *MiddleBlock =
1757 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1758 BasicBlock *ScalarPH =
1759 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1761 // Create and register the new vector loop.
1762 Loop* Lp = new Loop();
1763 Loop *ParentLoop = OrigLoop->getParentLoop();
1765 // Insert the new loop into the loop nest and register the new basic blocks
1766 // before calling any utilities such as SCEV that require valid LoopInfo.
1768 ParentLoop->addChildLoop(Lp);
1769 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1770 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1771 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1773 LI->addTopLevelLoop(Lp);
1775 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1777 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1779 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1781 // Generate the induction variable.
1782 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1783 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1784 // The loop step is equal to the vectorization factor (num of SIMD elements)
1785 // times the unroll factor (num of SIMD instructions).
1786 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1788 // This is the IR builder that we use to add all of the logic for bypassing
1789 // the new vector loop.
1790 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1791 setDebugLocFromInst(BypassBuilder,
1792 getDebugLocFromInstOrOperands(OldInduction));
1794 // We may need to extend the index in case there is a type mismatch.
1795 // We know that the count starts at zero and does not overflow.
1796 if (Count->getType() != IdxTy) {
1797 // The exit count can be of pointer type. Convert it to the correct
1799 if (ExitCount->getType()->isPointerTy())
1800 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1802 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1805 // Add the start index to the loop count to get the new end index.
1806 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1808 // Now we need to generate the expression for N - (N % VF), which is
1809 // the part that the vectorized body will execute.
1810 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1811 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1812 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1813 "end.idx.rnd.down");
1815 // Now, compare the new count to zero. If it is zero skip the vector loop and
1816 // jump to the scalar loop.
1817 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1820 BasicBlock *LastBypassBlock = BypassBlock;
1822 // Generate the code to check that the strides we assumed to be one are really
1823 // one. We want the new basic block to start at the first instruction in a
1824 // sequence of instructions that form a check.
1825 Instruction *StrideCheck;
1826 Instruction *FirstCheckInst;
1827 tie(FirstCheckInst, StrideCheck) =
1828 addStrideCheck(BypassBlock->getTerminator());
1830 // Create a new block containing the stride check.
1831 BasicBlock *CheckBlock =
1832 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
1834 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1835 LoopBypassBlocks.push_back(CheckBlock);
1837 // Replace the branch into the memory check block with a conditional branch
1838 // for the "few elements case".
1839 Instruction *OldTerm = BypassBlock->getTerminator();
1840 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1841 OldTerm->eraseFromParent();
1844 LastBypassBlock = CheckBlock;
1847 // Generate the code that checks in runtime if arrays overlap. We put the
1848 // checks into a separate block to make the more common case of few elements
1850 Instruction *MemRuntimeCheck;
1851 tie(FirstCheckInst, MemRuntimeCheck) =
1852 addRuntimeCheck(LastBypassBlock->getTerminator());
1853 if (MemRuntimeCheck) {
1854 // Create a new block containing the memory check.
1855 BasicBlock *CheckBlock =
1856 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
1858 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1859 LoopBypassBlocks.push_back(CheckBlock);
1861 // Replace the branch into the memory check block with a conditional branch
1862 // for the "few elements case".
1863 Instruction *OldTerm = LastBypassBlock->getTerminator();
1864 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1865 OldTerm->eraseFromParent();
1867 Cmp = MemRuntimeCheck;
1868 LastBypassBlock = CheckBlock;
1871 LastBypassBlock->getTerminator()->eraseFromParent();
1872 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1875 // We are going to resume the execution of the scalar loop.
1876 // Go over all of the induction variables that we found and fix the
1877 // PHIs that are left in the scalar version of the loop.
1878 // The starting values of PHI nodes depend on the counter of the last
1879 // iteration in the vectorized loop.
1880 // If we come from a bypass edge then we need to start from the original
1883 // This variable saves the new starting index for the scalar loop.
1884 PHINode *ResumeIndex = 0;
1885 LoopVectorizationLegality::InductionList::iterator I, E;
1886 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1887 // Set builder to point to last bypass block.
1888 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1889 for (I = List->begin(), E = List->end(); I != E; ++I) {
1890 PHINode *OrigPhi = I->first;
1891 LoopVectorizationLegality::InductionInfo II = I->second;
1893 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1894 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1895 MiddleBlock->getTerminator());
1896 // We might have extended the type of the induction variable but we need a
1897 // truncated version for the scalar loop.
1898 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1899 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1900 MiddleBlock->getTerminator()) : 0;
1902 Value *EndValue = 0;
1904 case LoopVectorizationLegality::IK_NoInduction:
1905 llvm_unreachable("Unknown induction");
1906 case LoopVectorizationLegality::IK_IntInduction: {
1907 // Handle the integer induction counter.
1908 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1910 // We have the canonical induction variable.
1911 if (OrigPhi == OldInduction) {
1912 // Create a truncated version of the resume value for the scalar loop,
1913 // we might have promoted the type to a larger width.
1915 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1916 // The new PHI merges the original incoming value, in case of a bypass,
1917 // or the value at the end of the vectorized loop.
1918 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1919 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1920 TruncResumeVal->addIncoming(EndValue, VecBody);
1922 // We know what the end value is.
1923 EndValue = IdxEndRoundDown;
1924 // We also know which PHI node holds it.
1925 ResumeIndex = ResumeVal;
1929 // Not the canonical induction variable - add the vector loop count to the
1931 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1932 II.StartValue->getType(),
1934 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1937 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1938 // Convert the CountRoundDown variable to the PHI size.
1939 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1940 II.StartValue->getType(),
1942 // Handle reverse integer induction counter.
1943 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1946 case LoopVectorizationLegality::IK_PtrInduction: {
1947 // For pointer induction variables, calculate the offset using
1949 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1953 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1954 // The value at the end of the loop for the reverse pointer is calculated
1955 // by creating a GEP with a negative index starting from the start value.
1956 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1957 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1959 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1965 // The new PHI merges the original incoming value, in case of a bypass,
1966 // or the value at the end of the vectorized loop.
1967 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1968 if (OrigPhi == OldInduction)
1969 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1971 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1973 ResumeVal->addIncoming(EndValue, VecBody);
1975 // Fix the scalar body counter (PHI node).
1976 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1977 // The old inductions phi node in the scalar body needs the truncated value.
1978 if (OrigPhi == OldInduction)
1979 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1981 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1984 // If we are generating a new induction variable then we also need to
1985 // generate the code that calculates the exit value. This value is not
1986 // simply the end of the counter because we may skip the vectorized body
1987 // in case of a runtime check.
1989 assert(!ResumeIndex && "Unexpected resume value found");
1990 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1991 MiddleBlock->getTerminator());
1992 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1993 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1994 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1997 // Make sure that we found the index where scalar loop needs to continue.
1998 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1999 "Invalid resume Index");
2001 // Add a check in the middle block to see if we have completed
2002 // all of the iterations in the first vector loop.
2003 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2004 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2005 ResumeIndex, "cmp.n",
2006 MiddleBlock->getTerminator());
2008 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2009 // Remove the old terminator.
2010 MiddleBlock->getTerminator()->eraseFromParent();
2012 // Create i+1 and fill the PHINode.
2013 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2014 Induction->addIncoming(StartIdx, VectorPH);
2015 Induction->addIncoming(NextIdx, VecBody);
2016 // Create the compare.
2017 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2018 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2020 // Now we have two terminators. Remove the old one from the block.
2021 VecBody->getTerminator()->eraseFromParent();
2023 // Get ready to start creating new instructions into the vectorized body.
2024 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2027 LoopVectorPreHeader = VectorPH;
2028 LoopScalarPreHeader = ScalarPH;
2029 LoopMiddleBlock = MiddleBlock;
2030 LoopExitBlock = ExitBlock;
2031 LoopVectorBody = VecBody;
2032 LoopScalarBody = OldBasicBlock;
2034 LoopVectorizeHints Hints(Lp, true);
2035 Hints.setAlreadyVectorized(Lp);
2038 /// This function returns the identity element (or neutral element) for
2039 /// the operation K.
2041 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2046 // Adding, Xoring, Oring zero to a number does not change it.
2047 return ConstantInt::get(Tp, 0);
2048 case RK_IntegerMult:
2049 // Multiplying a number by 1 does not change it.
2050 return ConstantInt::get(Tp, 1);
2052 // AND-ing a number with an all-1 value does not change it.
2053 return ConstantInt::get(Tp, -1, true);
2055 // Multiplying a number by 1 does not change it.
2056 return ConstantFP::get(Tp, 1.0L);
2058 // Adding zero to a number does not change it.
2059 return ConstantFP::get(Tp, 0.0L);
2061 llvm_unreachable("Unknown reduction kind");
2065 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2066 Intrinsic::ID ValidIntrinsicID) {
2067 if (I.getNumArgOperands() != 1 ||
2068 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2069 I.getType() != I.getArgOperand(0)->getType() ||
2070 !I.onlyReadsMemory())
2071 return Intrinsic::not_intrinsic;
2073 return ValidIntrinsicID;
2076 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2077 Intrinsic::ID ValidIntrinsicID) {
2078 if (I.getNumArgOperands() != 2 ||
2079 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2080 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2081 I.getType() != I.getArgOperand(0)->getType() ||
2082 I.getType() != I.getArgOperand(1)->getType() ||
2083 !I.onlyReadsMemory())
2084 return Intrinsic::not_intrinsic;
2086 return ValidIntrinsicID;
2090 static Intrinsic::ID
2091 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2092 // If we have an intrinsic call, check if it is trivially vectorizable.
2093 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2094 switch (II->getIntrinsicID()) {
2095 case Intrinsic::sqrt:
2096 case Intrinsic::sin:
2097 case Intrinsic::cos:
2098 case Intrinsic::exp:
2099 case Intrinsic::exp2:
2100 case Intrinsic::log:
2101 case Intrinsic::log10:
2102 case Intrinsic::log2:
2103 case Intrinsic::fabs:
2104 case Intrinsic::copysign:
2105 case Intrinsic::floor:
2106 case Intrinsic::ceil:
2107 case Intrinsic::trunc:
2108 case Intrinsic::rint:
2109 case Intrinsic::nearbyint:
2110 case Intrinsic::round:
2111 case Intrinsic::pow:
2112 case Intrinsic::fma:
2113 case Intrinsic::fmuladd:
2114 case Intrinsic::lifetime_start:
2115 case Intrinsic::lifetime_end:
2116 return II->getIntrinsicID();
2118 return Intrinsic::not_intrinsic;
2123 return Intrinsic::not_intrinsic;
2126 Function *F = CI->getCalledFunction();
2127 // We're going to make assumptions on the semantics of the functions, check
2128 // that the target knows that it's available in this environment and it does
2129 // not have local linkage.
2130 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2131 return Intrinsic::not_intrinsic;
2133 // Otherwise check if we have a call to a function that can be turned into a
2134 // vector intrinsic.
2141 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2145 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2149 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2151 case LibFunc::exp2f:
2152 case LibFunc::exp2l:
2153 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2157 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2158 case LibFunc::log10:
2159 case LibFunc::log10f:
2160 case LibFunc::log10l:
2161 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2163 case LibFunc::log2f:
2164 case LibFunc::log2l:
2165 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2167 case LibFunc::fabsf:
2168 case LibFunc::fabsl:
2169 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2170 case LibFunc::copysign:
2171 case LibFunc::copysignf:
2172 case LibFunc::copysignl:
2173 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2174 case LibFunc::floor:
2175 case LibFunc::floorf:
2176 case LibFunc::floorl:
2177 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2179 case LibFunc::ceilf:
2180 case LibFunc::ceill:
2181 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2182 case LibFunc::trunc:
2183 case LibFunc::truncf:
2184 case LibFunc::truncl:
2185 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2187 case LibFunc::rintf:
2188 case LibFunc::rintl:
2189 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2190 case LibFunc::nearbyint:
2191 case LibFunc::nearbyintf:
2192 case LibFunc::nearbyintl:
2193 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2194 case LibFunc::round:
2195 case LibFunc::roundf:
2196 case LibFunc::roundl:
2197 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2201 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2204 return Intrinsic::not_intrinsic;
2207 /// This function translates the reduction kind to an LLVM binary operator.
2209 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2211 case LoopVectorizationLegality::RK_IntegerAdd:
2212 return Instruction::Add;
2213 case LoopVectorizationLegality::RK_IntegerMult:
2214 return Instruction::Mul;
2215 case LoopVectorizationLegality::RK_IntegerOr:
2216 return Instruction::Or;
2217 case LoopVectorizationLegality::RK_IntegerAnd:
2218 return Instruction::And;
2219 case LoopVectorizationLegality::RK_IntegerXor:
2220 return Instruction::Xor;
2221 case LoopVectorizationLegality::RK_FloatMult:
2222 return Instruction::FMul;
2223 case LoopVectorizationLegality::RK_FloatAdd:
2224 return Instruction::FAdd;
2225 case LoopVectorizationLegality::RK_IntegerMinMax:
2226 return Instruction::ICmp;
2227 case LoopVectorizationLegality::RK_FloatMinMax:
2228 return Instruction::FCmp;
2230 llvm_unreachable("Unknown reduction operation");
2234 Value *createMinMaxOp(IRBuilder<> &Builder,
2235 LoopVectorizationLegality::MinMaxReductionKind RK,
2238 CmpInst::Predicate P = CmpInst::ICMP_NE;
2241 llvm_unreachable("Unknown min/max reduction kind");
2242 case LoopVectorizationLegality::MRK_UIntMin:
2243 P = CmpInst::ICMP_ULT;
2245 case LoopVectorizationLegality::MRK_UIntMax:
2246 P = CmpInst::ICMP_UGT;
2248 case LoopVectorizationLegality::MRK_SIntMin:
2249 P = CmpInst::ICMP_SLT;
2251 case LoopVectorizationLegality::MRK_SIntMax:
2252 P = CmpInst::ICMP_SGT;
2254 case LoopVectorizationLegality::MRK_FloatMin:
2255 P = CmpInst::FCMP_OLT;
2257 case LoopVectorizationLegality::MRK_FloatMax:
2258 P = CmpInst::FCMP_OGT;
2263 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2264 RK == LoopVectorizationLegality::MRK_FloatMax)
2265 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2267 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2269 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2274 struct CSEDenseMapInfo {
2275 static bool canHandle(Instruction *I) {
2276 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2277 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2279 static inline Instruction *getEmptyKey() {
2280 return DenseMapInfo<Instruction *>::getEmptyKey();
2282 static inline Instruction *getTombstoneKey() {
2283 return DenseMapInfo<Instruction *>::getTombstoneKey();
2285 static unsigned getHashValue(Instruction *I) {
2286 assert(canHandle(I) && "Unknown instruction!");
2287 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2288 I->value_op_end()));
2290 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2291 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2292 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2294 return LHS->isIdenticalTo(RHS);
2299 ///\brief Perform cse of induction variable instructions.
2300 static void cse(BasicBlock *BB) {
2301 // Perform simple cse.
2302 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2303 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2304 Instruction *In = I++;
2306 if (!CSEDenseMapInfo::canHandle(In))
2309 // Check if we can replace this instruction with any of the
2310 // visited instructions.
2311 if (Instruction *V = CSEMap.lookup(In)) {
2312 In->replaceAllUsesWith(V);
2313 In->eraseFromParent();
2321 void InnerLoopVectorizer::vectorizeLoop() {
2322 //===------------------------------------------------===//
2324 // Notice: any optimization or new instruction that go
2325 // into the code below should be also be implemented in
2328 //===------------------------------------------------===//
2329 Constant *Zero = Builder.getInt32(0);
2331 // In order to support reduction variables we need to be able to vectorize
2332 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2333 // stages. First, we create a new vector PHI node with no incoming edges.
2334 // We use this value when we vectorize all of the instructions that use the
2335 // PHI. Next, after all of the instructions in the block are complete we
2336 // add the new incoming edges to the PHI. At this point all of the
2337 // instructions in the basic block are vectorized, so we can use them to
2338 // construct the PHI.
2339 PhiVector RdxPHIsToFix;
2341 // Scan the loop in a topological order to ensure that defs are vectorized
2343 LoopBlocksDFS DFS(OrigLoop);
2346 // Vectorize all of the blocks in the original loop.
2347 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2348 be = DFS.endRPO(); bb != be; ++bb)
2349 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2351 // At this point every instruction in the original loop is widened to
2352 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2353 // that we vectorized. The PHI nodes are currently empty because we did
2354 // not want to introduce cycles. Notice that the remaining PHI nodes
2355 // that we need to fix are reduction variables.
2357 // Create the 'reduced' values for each of the induction vars.
2358 // The reduced values are the vector values that we scalarize and combine
2359 // after the loop is finished.
2360 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2362 PHINode *RdxPhi = *it;
2363 assert(RdxPhi && "Unable to recover vectorized PHI");
2365 // Find the reduction variable descriptor.
2366 assert(Legal->getReductionVars()->count(RdxPhi) &&
2367 "Unable to find the reduction variable");
2368 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2369 (*Legal->getReductionVars())[RdxPhi];
2371 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2373 // We need to generate a reduction vector from the incoming scalar.
2374 // To do so, we need to generate the 'identity' vector and override
2375 // one of the elements with the incoming scalar reduction. We need
2376 // to do it in the vector-loop preheader.
2377 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2379 // This is the vector-clone of the value that leaves the loop.
2380 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2381 Type *VecTy = VectorExit[0]->getType();
2383 // Find the reduction identity variable. Zero for addition, or, xor,
2384 // one for multiplication, -1 for And.
2387 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2388 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2389 // MinMax reduction have the start value as their identify.
2391 VectorStart = Identity = RdxDesc.StartValue;
2393 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2398 // Handle other reduction kinds:
2400 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2401 VecTy->getScalarType());
2404 // This vector is the Identity vector where the first element is the
2405 // incoming scalar reduction.
2406 VectorStart = RdxDesc.StartValue;
2408 Identity = ConstantVector::getSplat(VF, Iden);
2410 // This vector is the Identity vector where the first element is the
2411 // incoming scalar reduction.
2412 VectorStart = Builder.CreateInsertElement(Identity,
2413 RdxDesc.StartValue, Zero);
2417 // Fix the vector-loop phi.
2418 // We created the induction variable so we know that the
2419 // preheader is the first entry.
2420 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2422 // Reductions do not have to start at zero. They can start with
2423 // any loop invariant values.
2424 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2425 BasicBlock *Latch = OrigLoop->getLoopLatch();
2426 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2427 VectorParts &Val = getVectorValue(LoopVal);
2428 for (unsigned part = 0; part < UF; ++part) {
2429 // Make sure to add the reduction stat value only to the
2430 // first unroll part.
2431 Value *StartVal = (part == 0) ? VectorStart : Identity;
2432 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2433 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2436 // Before each round, move the insertion point right between
2437 // the PHIs and the values we are going to write.
2438 // This allows us to write both PHINodes and the extractelement
2440 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2442 VectorParts RdxParts;
2443 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2444 for (unsigned part = 0; part < UF; ++part) {
2445 // This PHINode contains the vectorized reduction variable, or
2446 // the initial value vector, if we bypass the vector loop.
2447 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2448 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2449 Value *StartVal = (part == 0) ? VectorStart : Identity;
2450 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2451 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2452 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2453 RdxParts.push_back(NewPhi);
2456 // Reduce all of the unrolled parts into a single vector.
2457 Value *ReducedPartRdx = RdxParts[0];
2458 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2459 setDebugLocFromInst(Builder, ReducedPartRdx);
2460 for (unsigned part = 1; part < UF; ++part) {
2461 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2462 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2463 RdxParts[part], ReducedPartRdx,
2466 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2467 ReducedPartRdx, RdxParts[part]);
2471 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2472 // and vector ops, reducing the set of values being computed by half each
2474 assert(isPowerOf2_32(VF) &&
2475 "Reduction emission only supported for pow2 vectors!");
2476 Value *TmpVec = ReducedPartRdx;
2477 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2478 for (unsigned i = VF; i != 1; i >>= 1) {
2479 // Move the upper half of the vector to the lower half.
2480 for (unsigned j = 0; j != i/2; ++j)
2481 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2483 // Fill the rest of the mask with undef.
2484 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2485 UndefValue::get(Builder.getInt32Ty()));
2488 Builder.CreateShuffleVector(TmpVec,
2489 UndefValue::get(TmpVec->getType()),
2490 ConstantVector::get(ShuffleMask),
2493 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2494 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2497 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2500 // The result is in the first element of the vector.
2501 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2502 Builder.getInt32(0));
2505 // Now, we need to fix the users of the reduction variable
2506 // inside and outside of the scalar remainder loop.
2507 // We know that the loop is in LCSSA form. We need to update the
2508 // PHI nodes in the exit blocks.
2509 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2510 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2511 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2512 if (!LCSSAPhi) break;
2514 // All PHINodes need to have a single entry edge, or two if
2515 // we already fixed them.
2516 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2518 // We found our reduction value exit-PHI. Update it with the
2519 // incoming bypass edge.
2520 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2521 // Add an edge coming from the bypass.
2522 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2525 }// end of the LCSSA phi scan.
2527 // Fix the scalar loop reduction variable with the incoming reduction sum
2528 // from the vector body and from the backedge value.
2529 int IncomingEdgeBlockIdx =
2530 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2531 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2532 // Pick the other block.
2533 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2534 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2535 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2536 }// end of for each redux variable.
2540 // Remove redundant induction instructions.
2541 cse(LoopVectorBody);
2544 void InnerLoopVectorizer::fixLCSSAPHIs() {
2545 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2546 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2547 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2548 if (!LCSSAPhi) break;
2549 if (LCSSAPhi->getNumIncomingValues() == 1)
2550 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2555 InnerLoopVectorizer::VectorParts
2556 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2557 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2560 // Look for cached value.
2561 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2562 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2563 if (ECEntryIt != MaskCache.end())
2564 return ECEntryIt->second;
2566 VectorParts SrcMask = createBlockInMask(Src);
2568 // The terminator has to be a branch inst!
2569 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2570 assert(BI && "Unexpected terminator found");
2572 if (BI->isConditional()) {
2573 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2575 if (BI->getSuccessor(0) != Dst)
2576 for (unsigned part = 0; part < UF; ++part)
2577 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2579 for (unsigned part = 0; part < UF; ++part)
2580 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2582 MaskCache[Edge] = EdgeMask;
2586 MaskCache[Edge] = SrcMask;
2590 InnerLoopVectorizer::VectorParts
2591 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2592 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2594 // Loop incoming mask is all-one.
2595 if (OrigLoop->getHeader() == BB) {
2596 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2597 return getVectorValue(C);
2600 // This is the block mask. We OR all incoming edges, and with zero.
2601 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2602 VectorParts BlockMask = getVectorValue(Zero);
2605 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2606 VectorParts EM = createEdgeMask(*it, BB);
2607 for (unsigned part = 0; part < UF; ++part)
2608 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2614 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2615 InnerLoopVectorizer::VectorParts &Entry,
2616 unsigned UF, unsigned VF, PhiVector *PV) {
2617 PHINode* P = cast<PHINode>(PN);
2618 // Handle reduction variables:
2619 if (Legal->getReductionVars()->count(P)) {
2620 for (unsigned part = 0; part < UF; ++part) {
2621 // This is phase one of vectorizing PHIs.
2622 Type *VecTy = (VF == 1) ? PN->getType() :
2623 VectorType::get(PN->getType(), VF);
2624 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2625 LoopVectorBody-> getFirstInsertionPt());
2631 setDebugLocFromInst(Builder, P);
2632 // Check for PHI nodes that are lowered to vector selects.
2633 if (P->getParent() != OrigLoop->getHeader()) {
2634 // We know that all PHIs in non-header blocks are converted into
2635 // selects, so we don't have to worry about the insertion order and we
2636 // can just use the builder.
2637 // At this point we generate the predication tree. There may be
2638 // duplications since this is a simple recursive scan, but future
2639 // optimizations will clean it up.
2641 unsigned NumIncoming = P->getNumIncomingValues();
2643 // Generate a sequence of selects of the form:
2644 // SELECT(Mask3, In3,
2645 // SELECT(Mask2, In2,
2647 for (unsigned In = 0; In < NumIncoming; In++) {
2648 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2650 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2652 for (unsigned part = 0; part < UF; ++part) {
2653 // We might have single edge PHIs (blocks) - use an identity
2654 // 'select' for the first PHI operand.
2656 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2659 // Select between the current value and the previous incoming edge
2660 // based on the incoming mask.
2661 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2662 Entry[part], "predphi");
2668 // This PHINode must be an induction variable.
2669 // Make sure that we know about it.
2670 assert(Legal->getInductionVars()->count(P) &&
2671 "Not an induction variable");
2673 LoopVectorizationLegality::InductionInfo II =
2674 Legal->getInductionVars()->lookup(P);
2677 case LoopVectorizationLegality::IK_NoInduction:
2678 llvm_unreachable("Unknown induction");
2679 case LoopVectorizationLegality::IK_IntInduction: {
2680 assert(P->getType() == II.StartValue->getType() && "Types must match");
2681 Type *PhiTy = P->getType();
2683 if (P == OldInduction) {
2684 // Handle the canonical induction variable. We might have had to
2686 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2688 // Handle other induction variables that are now based on the
2690 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2692 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2693 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2696 Broadcasted = getBroadcastInstrs(Broadcasted);
2697 // After broadcasting the induction variable we need to make the vector
2698 // consecutive by adding 0, 1, 2, etc.
2699 for (unsigned part = 0; part < UF; ++part)
2700 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2703 case LoopVectorizationLegality::IK_ReverseIntInduction:
2704 case LoopVectorizationLegality::IK_PtrInduction:
2705 case LoopVectorizationLegality::IK_ReversePtrInduction:
2706 // Handle reverse integer and pointer inductions.
2707 Value *StartIdx = ExtendedIdx;
2708 // This is the normalized GEP that starts counting at zero.
2709 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2712 // Handle the reverse integer induction variable case.
2713 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2714 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2715 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2717 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2720 // This is a new value so do not hoist it out.
2721 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2722 // After broadcasting the induction variable we need to make the
2723 // vector consecutive by adding ... -3, -2, -1, 0.
2724 for (unsigned part = 0; part < UF; ++part)
2725 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2730 // Handle the pointer induction variable case.
2731 assert(P->getType()->isPointerTy() && "Unexpected type.");
2733 // Is this a reverse induction ptr or a consecutive induction ptr.
2734 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2737 // This is the vector of results. Notice that we don't generate
2738 // vector geps because scalar geps result in better code.
2739 for (unsigned part = 0; part < UF; ++part) {
2741 int EltIndex = (part) * (Reverse ? -1 : 1);
2742 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2745 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2747 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2749 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2751 Entry[part] = SclrGep;
2755 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2756 for (unsigned int i = 0; i < VF; ++i) {
2757 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2758 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2761 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2763 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2765 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2767 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2768 Builder.getInt32(i),
2771 Entry[part] = VecVal;
2777 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2778 // For each instruction in the old loop.
2779 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2780 VectorParts &Entry = WidenMap.get(it);
2781 switch (it->getOpcode()) {
2782 case Instruction::Br:
2783 // Nothing to do for PHIs and BR, since we already took care of the
2784 // loop control flow instructions.
2786 case Instruction::PHI:{
2787 // Vectorize PHINodes.
2788 widenPHIInstruction(it, Entry, UF, VF, PV);
2792 case Instruction::Add:
2793 case Instruction::FAdd:
2794 case Instruction::Sub:
2795 case Instruction::FSub:
2796 case Instruction::Mul:
2797 case Instruction::FMul:
2798 case Instruction::UDiv:
2799 case Instruction::SDiv:
2800 case Instruction::FDiv:
2801 case Instruction::URem:
2802 case Instruction::SRem:
2803 case Instruction::FRem:
2804 case Instruction::Shl:
2805 case Instruction::LShr:
2806 case Instruction::AShr:
2807 case Instruction::And:
2808 case Instruction::Or:
2809 case Instruction::Xor: {
2810 // Just widen binops.
2811 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2812 setDebugLocFromInst(Builder, BinOp);
2813 VectorParts &A = getVectorValue(it->getOperand(0));
2814 VectorParts &B = getVectorValue(it->getOperand(1));
2816 // Use this vector value for all users of the original instruction.
2817 for (unsigned Part = 0; Part < UF; ++Part) {
2818 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2820 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2821 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2822 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2823 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2824 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2826 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2827 VecOp->setIsExact(BinOp->isExact());
2833 case Instruction::Select: {
2835 // If the selector is loop invariant we can create a select
2836 // instruction with a scalar condition. Otherwise, use vector-select.
2837 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2839 setDebugLocFromInst(Builder, it);
2841 // The condition can be loop invariant but still defined inside the
2842 // loop. This means that we can't just use the original 'cond' value.
2843 // We have to take the 'vectorized' value and pick the first lane.
2844 // Instcombine will make this a no-op.
2845 VectorParts &Cond = getVectorValue(it->getOperand(0));
2846 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2847 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2849 Value *ScalarCond = (VF == 1) ? Cond[0] :
2850 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2852 for (unsigned Part = 0; Part < UF; ++Part) {
2853 Entry[Part] = Builder.CreateSelect(
2854 InvariantCond ? ScalarCond : Cond[Part],
2861 case Instruction::ICmp:
2862 case Instruction::FCmp: {
2863 // Widen compares. Generate vector compares.
2864 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2865 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2866 setDebugLocFromInst(Builder, it);
2867 VectorParts &A = getVectorValue(it->getOperand(0));
2868 VectorParts &B = getVectorValue(it->getOperand(1));
2869 for (unsigned Part = 0; Part < UF; ++Part) {
2872 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2874 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2880 case Instruction::Store:
2881 case Instruction::Load:
2882 vectorizeMemoryInstruction(it);
2884 case Instruction::ZExt:
2885 case Instruction::SExt:
2886 case Instruction::FPToUI:
2887 case Instruction::FPToSI:
2888 case Instruction::FPExt:
2889 case Instruction::PtrToInt:
2890 case Instruction::IntToPtr:
2891 case Instruction::SIToFP:
2892 case Instruction::UIToFP:
2893 case Instruction::Trunc:
2894 case Instruction::FPTrunc:
2895 case Instruction::BitCast: {
2896 CastInst *CI = dyn_cast<CastInst>(it);
2897 setDebugLocFromInst(Builder, it);
2898 /// Optimize the special case where the source is the induction
2899 /// variable. Notice that we can only optimize the 'trunc' case
2900 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2901 /// c. other casts depend on pointer size.
2902 if (CI->getOperand(0) == OldInduction &&
2903 it->getOpcode() == Instruction::Trunc) {
2904 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2906 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2907 for (unsigned Part = 0; Part < UF; ++Part)
2908 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2911 /// Vectorize casts.
2912 Type *DestTy = (VF == 1) ? CI->getType() :
2913 VectorType::get(CI->getType(), VF);
2915 VectorParts &A = getVectorValue(it->getOperand(0));
2916 for (unsigned Part = 0; Part < UF; ++Part)
2917 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2921 case Instruction::Call: {
2922 // Ignore dbg intrinsics.
2923 if (isa<DbgInfoIntrinsic>(it))
2925 setDebugLocFromInst(Builder, it);
2927 Module *M = BB->getParent()->getParent();
2928 CallInst *CI = cast<CallInst>(it);
2929 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2930 assert(ID && "Not an intrinsic call!");
2932 case Intrinsic::lifetime_end:
2933 case Intrinsic::lifetime_start:
2934 scalarizeInstruction(it);
2937 for (unsigned Part = 0; Part < UF; ++Part) {
2938 SmallVector<Value *, 4> Args;
2939 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2940 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2941 Args.push_back(Arg[Part]);
2943 Type *Tys[] = {CI->getType()};
2945 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2947 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2948 Entry[Part] = Builder.CreateCall(F, Args);
2956 // All other instructions are unsupported. Scalarize them.
2957 scalarizeInstruction(it);
2960 }// end of for_each instr.
2963 void InnerLoopVectorizer::updateAnalysis() {
2964 // Forget the original basic block.
2965 SE->forgetLoop(OrigLoop);
2967 // Update the dominator tree information.
2968 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2969 "Entry does not dominate exit.");
2971 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2972 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2973 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2974 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2975 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2976 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2977 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2978 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2980 DEBUG(DT->verifyDomTree());
2983 /// \brief Check whether it is safe to if-convert this phi node.
2985 /// Phi nodes with constant expressions that can trap are not safe to if
2987 static bool canIfConvertPHINodes(BasicBlock *BB) {
2988 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2989 PHINode *Phi = dyn_cast<PHINode>(I);
2992 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
2993 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3000 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3001 if (!EnableIfConversion)
3004 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3006 // A list of pointers that we can safely read and write to.
3007 SmallPtrSet<Value *, 8> SafePointes;
3009 // Collect safe addresses.
3010 for (Loop::block_iterator BI = TheLoop->block_begin(),
3011 BE = TheLoop->block_end(); BI != BE; ++BI) {
3012 BasicBlock *BB = *BI;
3014 if (blockNeedsPredication(BB))
3017 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3018 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3019 SafePointes.insert(LI->getPointerOperand());
3020 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3021 SafePointes.insert(SI->getPointerOperand());
3025 // Collect the blocks that need predication.
3026 BasicBlock *Header = TheLoop->getHeader();
3027 for (Loop::block_iterator BI = TheLoop->block_begin(),
3028 BE = TheLoop->block_end(); BI != BE; ++BI) {
3029 BasicBlock *BB = *BI;
3031 // We don't support switch statements inside loops.
3032 if (!isa<BranchInst>(BB->getTerminator()))
3035 // We must be able to predicate all blocks that need to be predicated.
3036 if (blockNeedsPredication(BB)) {
3037 if (!blockCanBePredicated(BB, SafePointes))
3039 } else if (BB != Header && !canIfConvertPHINodes(BB))
3044 // We can if-convert this loop.
3048 bool LoopVectorizationLegality::canVectorize() {
3049 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3050 // be canonicalized.
3051 if (!TheLoop->getLoopPreheader())
3054 // We can only vectorize innermost loops.
3055 if (TheLoop->getSubLoopsVector().size())
3058 // We must have a single backedge.
3059 if (TheLoop->getNumBackEdges() != 1)
3062 // We must have a single exiting block.
3063 if (!TheLoop->getExitingBlock())
3066 // We need to have a loop header.
3067 DEBUG(dbgs() << "LV: Found a loop: " <<
3068 TheLoop->getHeader()->getName() << '\n');
3070 // Check if we can if-convert non-single-bb loops.
3071 unsigned NumBlocks = TheLoop->getNumBlocks();
3072 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3073 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3077 // ScalarEvolution needs to be able to find the exit count.
3078 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3079 if (ExitCount == SE->getCouldNotCompute()) {
3080 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3084 // Do not loop-vectorize loops with a tiny trip count.
3085 BasicBlock *Latch = TheLoop->getLoopLatch();
3086 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
3087 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
3088 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
3089 "This loop is not worth vectorizing.\n");
3093 // Check if we can vectorize the instructions and CFG in this loop.
3094 if (!canVectorizeInstrs()) {
3095 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3099 // Go over each instruction and look at memory deps.
3100 if (!canVectorizeMemory()) {
3101 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3105 // Collect all of the variables that remain uniform after vectorization.
3106 collectLoopUniforms();
3108 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3109 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3112 // Okay! We can vectorize. At this point we don't have any other mem analysis
3113 // which may limit our maximum vectorization factor, so just return true with
3118 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
3119 if (Ty->isPointerTy())
3120 return DL.getIntPtrType(Ty);
3122 // It is possible that char's or short's overflow when we ask for the loop's
3123 // trip count, work around this by changing the type size.
3124 if (Ty->getScalarSizeInBits() < 32)
3125 return Type::getInt32Ty(Ty->getContext());
3130 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
3131 Ty0 = convertPointerToIntegerType(DL, Ty0);
3132 Ty1 = convertPointerToIntegerType(DL, Ty1);
3133 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3138 /// \brief Check that the instruction has outside loop users and is not an
3139 /// identified reduction variable.
3140 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3141 SmallPtrSet<Value *, 4> &Reductions) {
3142 // Reduction instructions are allowed to have exit users. All other
3143 // instructions must not have external users.
3144 if (!Reductions.count(Inst))
3145 //Check that all of the users of the loop are inside the BB.
3146 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
3148 Instruction *U = cast<Instruction>(*I);
3149 // This user may be a reduction exit value.
3150 if (!TheLoop->contains(U)) {
3151 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
3158 bool LoopVectorizationLegality::canVectorizeInstrs() {
3159 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3160 BasicBlock *Header = TheLoop->getHeader();
3162 // Look for the attribute signaling the absence of NaNs.
3163 Function &F = *Header->getParent();
3164 if (F.hasFnAttribute("no-nans-fp-math"))
3165 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3166 AttributeSet::FunctionIndex,
3167 "no-nans-fp-math").getValueAsString() == "true";
3169 // For each block in the loop.
3170 for (Loop::block_iterator bb = TheLoop->block_begin(),
3171 be = TheLoop->block_end(); bb != be; ++bb) {
3173 // Scan the instructions in the block and look for hazards.
3174 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3177 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3178 Type *PhiTy = Phi->getType();
3179 // Check that this PHI type is allowed.
3180 if (!PhiTy->isIntegerTy() &&
3181 !PhiTy->isFloatingPointTy() &&
3182 !PhiTy->isPointerTy()) {
3183 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3187 // If this PHINode is not in the header block, then we know that we
3188 // can convert it to select during if-conversion. No need to check if
3189 // the PHIs in this block are induction or reduction variables.
3190 if (*bb != Header) {
3191 // Check that this instruction has no outside users or is an
3192 // identified reduction value with an outside user.
3193 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3198 // We only allow if-converted PHIs with more than two incoming values.
3199 if (Phi->getNumIncomingValues() != 2) {
3200 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3204 // This is the value coming from the preheader.
3205 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3206 // Check if this is an induction variable.
3207 InductionKind IK = isInductionVariable(Phi);
3209 if (IK_NoInduction != IK) {
3210 // Get the widest type.
3212 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3214 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3216 // Int inductions are special because we only allow one IV.
3217 if (IK == IK_IntInduction) {
3218 // Use the phi node with the widest type as induction. Use the last
3219 // one if there are multiple (no good reason for doing this other
3220 // than it is expedient).
3221 if (!Induction || PhiTy == WidestIndTy)
3225 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3226 Inductions[Phi] = InductionInfo(StartValue, IK);
3228 // Until we explicitly handle the case of an induction variable with
3229 // an outside loop user we have to give up vectorizing this loop.
3230 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3236 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3237 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3240 if (AddReductionVar(Phi, RK_IntegerMult)) {
3241 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3244 if (AddReductionVar(Phi, RK_IntegerOr)) {
3245 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3248 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3249 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3252 if (AddReductionVar(Phi, RK_IntegerXor)) {
3253 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3256 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3257 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3260 if (AddReductionVar(Phi, RK_FloatMult)) {
3261 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3264 if (AddReductionVar(Phi, RK_FloatAdd)) {
3265 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3268 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3269 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3274 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3276 }// end of PHI handling
3278 // We still don't handle functions. However, we can ignore dbg intrinsic
3279 // calls and we do handle certain intrinsic and libm functions.
3280 CallInst *CI = dyn_cast<CallInst>(it);
3281 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3282 DEBUG(dbgs() << "LV: Found a call site.\n");
3286 // Check that the instruction return type is vectorizable.
3287 // Also, we can't vectorize extractelement instructions.
3288 if ((!VectorType::isValidElementType(it->getType()) &&
3289 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3290 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3294 // Check that the stored type is vectorizable.
3295 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3296 Type *T = ST->getValueOperand()->getType();
3297 if (!VectorType::isValidElementType(T))
3299 if (EnableMemAccessVersioning)
3300 collectStridedAcccess(ST);
3303 if (EnableMemAccessVersioning)
3304 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3305 collectStridedAcccess(LI);
3307 // Reduction instructions are allowed to have exit users.
3308 // All other instructions must not have external users.
3309 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3317 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3318 if (Inductions.empty())
3325 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3326 /// return the induction operand of the gep pointer.
3327 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3328 DataLayout *DL, Loop *Lp) {
3329 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3333 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3335 // Check that all of the gep indices are uniform except for our induction
3337 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3338 if (i != InductionOperand &&
3339 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3341 return GEP->getOperand(InductionOperand);
3344 ///\brief Look for a cast use of the passed value.
3345 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3346 Value *UniqueCast = 0;
3347 for (Value::use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end(); UI != UE;
3349 CastInst *CI = dyn_cast<CastInst>(*UI);
3350 if (CI && CI->getType() == Ty) {
3360 ///\brief Get the stride of a pointer access in a loop.
3361 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3362 /// pointer to the Value, or null otherwise.
3363 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3364 DataLayout *DL, Loop *Lp) {
3365 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3366 if (!PtrTy || PtrTy->isAggregateType())
3369 // Try to remove a gep instruction to make the pointer (actually index at this
3370 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3371 // pointer, otherwise, we are analyzing the index.
3372 Value *OrigPtr = Ptr;
3374 // The size of the pointer access.
3375 int64_t PtrAccessSize = 1;
3377 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3378 const SCEV *V = SE->getSCEV(Ptr);
3382 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3383 V = C->getOperand();
3385 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3389 V = S->getStepRecurrence(*SE);
3393 // Strip off the size of access multiplication if we are still analyzing the
3395 if (OrigPtr == Ptr) {
3396 DL->getTypeAllocSize(PtrTy->getElementType());
3397 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3398 if (M->getOperand(0)->getSCEVType() != scConstant)
3401 const APInt &APStepVal =
3402 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3404 // Huge step value - give up.
3405 if (APStepVal.getBitWidth() > 64)
3408 int64_t StepVal = APStepVal.getSExtValue();
3409 if (PtrAccessSize != StepVal)
3411 V = M->getOperand(1);
3416 Type *StripedOffRecurrenceCast = 0;
3417 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3418 StripedOffRecurrenceCast = C->getType();
3419 V = C->getOperand();
3422 // Look for the loop invariant symbolic value.
3423 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3427 Value *Stride = U->getValue();
3428 if (!Lp->isLoopInvariant(Stride))
3431 // If we have stripped off the recurrence cast we have to make sure that we
3432 // return the value that is used in this loop so that we can replace it later.
3433 if (StripedOffRecurrenceCast)
3434 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3439 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3441 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3442 Ptr = LI->getPointerOperand();
3443 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3444 Ptr = SI->getPointerOperand();
3448 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3452 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3453 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3454 Strides[Ptr] = Stride;
3455 StrideSet.insert(Stride);
3458 void LoopVectorizationLegality::collectLoopUniforms() {
3459 // We now know that the loop is vectorizable!
3460 // Collect variables that will remain uniform after vectorization.
3461 std::vector<Value*> Worklist;
3462 BasicBlock *Latch = TheLoop->getLoopLatch();
3464 // Start with the conditional branch and walk up the block.
3465 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3467 while (Worklist.size()) {
3468 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3469 Worklist.pop_back();
3471 // Look at instructions inside this loop.
3472 // Stop when reaching PHI nodes.
3473 // TODO: we need to follow values all over the loop, not only in this block.
3474 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3477 // This is a known uniform.
3480 // Insert all operands.
3481 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3486 /// \brief Analyses memory accesses in a loop.
3488 /// Checks whether run time pointer checks are needed and builds sets for data
3489 /// dependence checking.
3490 class AccessAnalysis {
3492 /// \brief Read or write access location.
3493 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3494 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3496 /// \brief Set of potential dependent memory accesses.
3497 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3499 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3500 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3501 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3503 /// \brief Register a load and whether it is only read from.
3504 void addLoad(Value *Ptr, bool IsReadOnly) {
3505 Accesses.insert(MemAccessInfo(Ptr, false));
3507 ReadOnlyPtr.insert(Ptr);
3510 /// \brief Register a store.
3511 void addStore(Value *Ptr) {
3512 Accesses.insert(MemAccessInfo(Ptr, true));
3515 /// \brief Check whether we can check the pointers at runtime for
3516 /// non-intersection.
3517 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3518 unsigned &NumComparisons, ScalarEvolution *SE,
3519 Loop *TheLoop, ValueToValueMap &Strides,
3520 bool ShouldCheckStride = false);
3522 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3523 /// and builds sets of dependent accesses.
3524 void buildDependenceSets() {
3525 // Process read-write pointers first.
3526 processMemAccesses(false);
3527 // Next, process read pointers.
3528 processMemAccesses(true);
3531 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3533 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3534 void resetDepChecks() { CheckDeps.clear(); }
3536 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3539 typedef SetVector<MemAccessInfo> PtrAccessSet;
3540 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3542 /// \brief Go over all memory access or only the deferred ones if
3543 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3544 /// and build sets of dependency check candidates.
3545 void processMemAccesses(bool UseDeferred);
3547 /// Set of all accesses.
3548 PtrAccessSet Accesses;
3550 /// Set of access to check after all writes have been processed.
3551 PtrAccessSet DeferredAccesses;
3553 /// Map of pointers to last access encountered.
3554 UnderlyingObjToAccessMap ObjToLastAccess;
3556 /// Set of accesses that need a further dependence check.
3557 MemAccessInfoSet CheckDeps;
3559 /// Set of pointers that are read only.
3560 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3562 /// Set of underlying objects already written to.
3563 SmallPtrSet<Value*, 16> WriteObjects;
3567 /// Sets of potentially dependent accesses - members of one set share an
3568 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3569 /// dependence check.
3570 DepCandidates &DepCands;
3572 bool AreAllWritesIdentified;
3573 bool AreAllReadsIdentified;
3574 bool IsRTCheckNeeded;
3577 } // end anonymous namespace
3579 /// \brief Check whether a pointer can participate in a runtime bounds check.
3580 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3582 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3583 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3587 return AR->isAffine();
3590 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3591 /// the address space.
3592 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3593 const Loop *Lp, ValueToValueMap &StridesMap);
3595 bool AccessAnalysis::canCheckPtrAtRT(
3596 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3597 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3598 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3599 // Find pointers with computable bounds. We are going to use this information
3600 // to place a runtime bound check.
3601 unsigned NumReadPtrChecks = 0;
3602 unsigned NumWritePtrChecks = 0;
3603 bool CanDoRT = true;
3605 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3606 // We assign consecutive id to access from different dependence sets.
3607 // Accesses within the same set don't need a runtime check.
3608 unsigned RunningDepId = 1;
3609 DenseMap<Value *, unsigned> DepSetId;
3611 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3613 const MemAccessInfo &Access = *AI;
3614 Value *Ptr = Access.getPointer();
3615 bool IsWrite = Access.getInt();
3617 // Just add write checks if we have both.
3618 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3622 ++NumWritePtrChecks;
3626 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3627 // When we run after a failing dependency check we have to make sure we
3628 // don't have wrapping pointers.
3629 (!ShouldCheckStride ||
3630 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3631 // The id of the dependence set.
3634 if (IsDepCheckNeeded) {
3635 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3636 unsigned &LeaderId = DepSetId[Leader];
3638 LeaderId = RunningDepId++;
3641 // Each access has its own dependence set.
3642 DepId = RunningDepId++;
3644 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3646 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3652 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3653 NumComparisons = 0; // Only one dependence set.
3655 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3656 NumWritePtrChecks - 1));
3659 // If the pointers that we would use for the bounds comparison have different
3660 // address spaces, assume the values aren't directly comparable, so we can't
3661 // use them for the runtime check. We also have to assume they could
3662 // overlap. In the future there should be metadata for whether address spaces
3664 unsigned NumPointers = RtCheck.Pointers.size();
3665 for (unsigned i = 0; i < NumPointers; ++i) {
3666 for (unsigned j = i + 1; j < NumPointers; ++j) {
3667 // Only need to check pointers between two different dependency sets.
3668 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3671 Value *PtrI = RtCheck.Pointers[i];
3672 Value *PtrJ = RtCheck.Pointers[j];
3674 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3675 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3677 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3678 " different address spaces\n");
3687 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3688 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3691 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3692 // We process the set twice: first we process read-write pointers, last we
3693 // process read-only pointers. This allows us to skip dependence tests for
3694 // read-only pointers.
3696 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3697 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3698 const MemAccessInfo &Access = *AI;
3699 Value *Ptr = Access.getPointer();
3700 bool IsWrite = Access.getInt();
3702 DepCands.insert(Access);
3704 // Memorize read-only pointers for later processing and skip them in the
3705 // first round (they need to be checked after we have seen all write
3706 // pointers). Note: we also mark pointer that are not consecutive as
3707 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3708 // second check for "!IsWrite".
3709 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3710 if (!UseDeferred && IsReadOnlyPtr) {
3711 DeferredAccesses.insert(Access);
3715 bool NeedDepCheck = false;
3716 // Check whether there is the possibility of dependency because of
3717 // underlying objects being the same.
3718 typedef SmallVector<Value*, 16> ValueVector;
3719 ValueVector TempObjects;
3720 GetUnderlyingObjects(Ptr, TempObjects, DL);
3721 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3723 Value *UnderlyingObj = *UI;
3725 // If this is a write then it needs to be an identified object. If this a
3726 // read and all writes (so far) are identified function scope objects we
3727 // don't need an identified underlying object but only an Argument (the
3728 // next write is going to invalidate this assumption if it is
3730 // This is a micro-optimization for the case where all writes are
3731 // identified and we have one argument pointer.
3732 // Otherwise, we do need a runtime check.
3733 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3734 (!IsWrite && (!AreAllWritesIdentified ||
3735 !isa<Argument>(UnderlyingObj)) &&
3736 !isIdentifiedObject(UnderlyingObj))) {
3737 DEBUG(dbgs() << "LV: Found an unidentified " <<
3738 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3740 IsRTCheckNeeded = (IsRTCheckNeeded ||
3741 !isIdentifiedObject(UnderlyingObj) ||
3742 !AreAllReadsIdentified);
3745 AreAllWritesIdentified = false;
3747 AreAllReadsIdentified = false;
3750 // If this is a write - check other reads and writes for conflicts. If
3751 // this is a read only check other writes for conflicts (but only if there
3752 // is no other write to the ptr - this is an optimization to catch "a[i] =
3753 // a[i] + " without having to do a dependence check).
3754 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3755 NeedDepCheck = true;
3758 WriteObjects.insert(UnderlyingObj);
3760 // Create sets of pointers connected by shared underlying objects.
3761 UnderlyingObjToAccessMap::iterator Prev =
3762 ObjToLastAccess.find(UnderlyingObj);
3763 if (Prev != ObjToLastAccess.end())
3764 DepCands.unionSets(Access, Prev->second);
3766 ObjToLastAccess[UnderlyingObj] = Access;
3770 CheckDeps.insert(Access);
3775 /// \brief Checks memory dependences among accesses to the same underlying
3776 /// object to determine whether there vectorization is legal or not (and at
3777 /// which vectorization factor).
3779 /// This class works under the assumption that we already checked that memory
3780 /// locations with different underlying pointers are "must-not alias".
3781 /// We use the ScalarEvolution framework to symbolically evalutate access
3782 /// functions pairs. Since we currently don't restructure the loop we can rely
3783 /// on the program order of memory accesses to determine their safety.
3784 /// At the moment we will only deem accesses as safe for:
3785 /// * A negative constant distance assuming program order.
3787 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3788 /// a[i] = tmp; y = a[i];
3790 /// The latter case is safe because later checks guarantuee that there can't
3791 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3792 /// the same variable: a header phi can only be an induction or a reduction, a
3793 /// reduction can't have a memory sink, an induction can't have a memory
3794 /// source). This is important and must not be violated (or we have to
3795 /// resort to checking for cycles through memory).
3797 /// * A positive constant distance assuming program order that is bigger
3798 /// than the biggest memory access.
3800 /// tmp = a[i] OR b[i] = x
3801 /// a[i+2] = tmp y = b[i+2];
3803 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3805 /// * Zero distances and all accesses have the same size.
3807 class MemoryDepChecker {
3809 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3810 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3812 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3813 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3814 ShouldRetryWithRuntimeCheck(false) {}
3816 /// \brief Register the location (instructions are given increasing numbers)
3817 /// of a write access.
3818 void addAccess(StoreInst *SI) {
3819 Value *Ptr = SI->getPointerOperand();
3820 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3821 InstMap.push_back(SI);
3825 /// \brief Register the location (instructions are given increasing numbers)
3826 /// of a write access.
3827 void addAccess(LoadInst *LI) {
3828 Value *Ptr = LI->getPointerOperand();
3829 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3830 InstMap.push_back(LI);
3834 /// \brief Check whether the dependencies between the accesses are safe.
3836 /// Only checks sets with elements in \p CheckDeps.
3837 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3838 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
3840 /// \brief The maximum number of bytes of a vector register we can vectorize
3841 /// the accesses safely with.
3842 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3844 /// \brief In same cases when the dependency check fails we can still
3845 /// vectorize the loop with a dynamic array access check.
3846 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3849 ScalarEvolution *SE;
3851 const Loop *InnermostLoop;
3853 /// \brief Maps access locations (ptr, read/write) to program order.
3854 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3856 /// \brief Memory access instructions in program order.
3857 SmallVector<Instruction *, 16> InstMap;
3859 /// \brief The program order index to be used for the next instruction.
3862 // We can access this many bytes in parallel safely.
3863 unsigned MaxSafeDepDistBytes;
3865 /// \brief If we see a non-constant dependence distance we can still try to
3866 /// vectorize this loop with runtime checks.
3867 bool ShouldRetryWithRuntimeCheck;
3869 /// \brief Check whether there is a plausible dependence between the two
3872 /// Access \p A must happen before \p B in program order. The two indices
3873 /// identify the index into the program order map.
3875 /// This function checks whether there is a plausible dependence (or the
3876 /// absence of such can't be proved) between the two accesses. If there is a
3877 /// plausible dependence but the dependence distance is bigger than one
3878 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3879 /// distance is smaller than any other distance encountered so far).
3880 /// Otherwise, this function returns true signaling a possible dependence.
3881 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3882 const MemAccessInfo &B, unsigned BIdx,
3883 ValueToValueMap &Strides);
3885 /// \brief Check whether the data dependence could prevent store-load
3887 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3890 } // end anonymous namespace
3892 static bool isInBoundsGep(Value *Ptr) {
3893 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3894 return GEP->isInBounds();
3898 /// \brief Check whether the access through \p Ptr has a constant stride.
3899 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3900 const Loop *Lp, ValueToValueMap &StridesMap) {
3901 const Type *Ty = Ptr->getType();
3902 assert(Ty->isPointerTy() && "Unexpected non-ptr");
3904 // Make sure that the pointer does not point to aggregate types.
3905 const PointerType *PtrTy = cast<PointerType>(Ty);
3906 if (PtrTy->getElementType()->isAggregateType()) {
3907 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3912 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
3914 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3916 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3917 << *Ptr << " SCEV: " << *PtrScev << "\n");
3921 // The accesss function must stride over the innermost loop.
3922 if (Lp != AR->getLoop()) {
3923 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3924 *Ptr << " SCEV: " << *PtrScev << "\n");
3927 // The address calculation must not wrap. Otherwise, a dependence could be
3929 // An inbounds getelementptr that is a AddRec with a unit stride
3930 // cannot wrap per definition. The unit stride requirement is checked later.
3931 // An getelementptr without an inbounds attribute and unit stride would have
3932 // to access the pointer value "0" which is undefined behavior in address
3933 // space 0, therefore we can also vectorize this case.
3934 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3935 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3936 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3937 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3938 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3939 << *Ptr << " SCEV: " << *PtrScev << "\n");
3943 // Check the step is constant.
3944 const SCEV *Step = AR->getStepRecurrence(*SE);
3946 // Calculate the pointer stride and check if it is consecutive.
3947 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3949 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3950 " SCEV: " << *PtrScev << "\n");
3954 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3955 const APInt &APStepVal = C->getValue()->getValue();
3957 // Huge step value - give up.
3958 if (APStepVal.getBitWidth() > 64)
3961 int64_t StepVal = APStepVal.getSExtValue();
3964 int64_t Stride = StepVal / Size;
3965 int64_t Rem = StepVal % Size;
3969 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3970 // know we can't "wrap around the address space". In case of address space
3971 // zero we know that this won't happen without triggering undefined behavior.
3972 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3973 Stride != 1 && Stride != -1)
3979 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3980 unsigned TypeByteSize) {
3981 // If loads occur at a distance that is not a multiple of a feasible vector
3982 // factor store-load forwarding does not take place.
3983 // Positive dependences might cause troubles because vectorizing them might
3984 // prevent store-load forwarding making vectorized code run a lot slower.
3985 // a[i] = a[i-3] ^ a[i-8];
3986 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3987 // hence on your typical architecture store-load forwarding does not take
3988 // place. Vectorizing in such cases does not make sense.
3989 // Store-load forwarding distance.
3990 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3991 // Maximum vector factor.
3992 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3993 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3994 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3996 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3998 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3999 MaxVFWithoutSLForwardIssues = (vf >>=1);
4004 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4005 DEBUG(dbgs() << "LV: Distance " << Distance <<
4006 " that could cause a store-load forwarding conflict\n");
4010 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4011 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4012 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4016 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4017 const MemAccessInfo &B, unsigned BIdx,
4018 ValueToValueMap &Strides) {
4019 assert (AIdx < BIdx && "Must pass arguments in program order");
4021 Value *APtr = A.getPointer();
4022 Value *BPtr = B.getPointer();
4023 bool AIsWrite = A.getInt();
4024 bool BIsWrite = B.getInt();
4026 // Two reads are independent.
4027 if (!AIsWrite && !BIsWrite)
4030 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4031 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4033 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4034 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4036 const SCEV *Src = AScev;
4037 const SCEV *Sink = BScev;
4039 // If the induction step is negative we have to invert source and sink of the
4041 if (StrideAPtr < 0) {
4044 std::swap(APtr, BPtr);
4045 std::swap(Src, Sink);
4046 std::swap(AIsWrite, BIsWrite);
4047 std::swap(AIdx, BIdx);
4048 std::swap(StrideAPtr, StrideBPtr);
4051 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4053 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4054 << "(Induction step: " << StrideAPtr << ")\n");
4055 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4056 << *InstMap[BIdx] << ": " << *Dist << "\n");
4058 // Need consecutive accesses. We don't want to vectorize
4059 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4060 // the address space.
4061 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4062 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4066 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4068 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4069 ShouldRetryWithRuntimeCheck = true;
4073 Type *ATy = APtr->getType()->getPointerElementType();
4074 Type *BTy = BPtr->getType()->getPointerElementType();
4075 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4077 // Negative distances are not plausible dependencies.
4078 const APInt &Val = C->getValue()->getValue();
4079 if (Val.isNegative()) {
4080 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4081 if (IsTrueDataDependence &&
4082 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4086 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4090 // Write to the same location with the same size.
4091 // Could be improved to assert type sizes are the same (i32 == float, etc).
4095 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4099 assert(Val.isStrictlyPositive() && "Expect a positive value");
4101 // Positive distance bigger than max vectorization factor.
4104 "LV: ReadWrite-Write positive dependency with different types\n");
4108 unsigned Distance = (unsigned) Val.getZExtValue();
4110 // Bail out early if passed-in parameters make vectorization not feasible.
4111 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4112 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4114 // The distance must be bigger than the size needed for a vectorized version
4115 // of the operation and the size of the vectorized operation must not be
4116 // bigger than the currrent maximum size.
4117 if (Distance < 2*TypeByteSize ||
4118 2*TypeByteSize > MaxSafeDepDistBytes ||
4119 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4120 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4121 << Val.getSExtValue() << '\n');
4125 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4126 Distance : MaxSafeDepDistBytes;
4128 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4129 if (IsTrueDataDependence &&
4130 couldPreventStoreLoadForward(Distance, TypeByteSize))
4133 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4134 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4139 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4140 MemAccessInfoSet &CheckDeps,
4141 ValueToValueMap &Strides) {
4143 MaxSafeDepDistBytes = -1U;
4144 while (!CheckDeps.empty()) {
4145 MemAccessInfo CurAccess = *CheckDeps.begin();
4147 // Get the relevant memory access set.
4148 EquivalenceClasses<MemAccessInfo>::iterator I =
4149 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4151 // Check accesses within this set.
4152 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4153 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4155 // Check every access pair.
4157 CheckDeps.erase(*AI);
4158 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
4160 // Check every accessing instruction pair in program order.
4161 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4162 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4163 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4164 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4165 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4167 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4178 bool LoopVectorizationLegality::canVectorizeMemory() {
4180 typedef SmallVector<Value*, 16> ValueVector;
4181 typedef SmallPtrSet<Value*, 16> ValueSet;
4183 // Holds the Load and Store *instructions*.
4187 // Holds all the different accesses in the loop.
4188 unsigned NumReads = 0;
4189 unsigned NumReadWrites = 0;
4191 PtrRtCheck.Pointers.clear();
4192 PtrRtCheck.Need = false;
4194 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4195 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4198 for (Loop::block_iterator bb = TheLoop->block_begin(),
4199 be = TheLoop->block_end(); bb != be; ++bb) {
4201 // Scan the BB and collect legal loads and stores.
4202 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4205 // If this is a load, save it. If this instruction can read from memory
4206 // but is not a load, then we quit. Notice that we don't handle function
4207 // calls that read or write.
4208 if (it->mayReadFromMemory()) {
4209 // Many math library functions read the rounding mode. We will only
4210 // vectorize a loop if it contains known function calls that don't set
4211 // the flag. Therefore, it is safe to ignore this read from memory.
4212 CallInst *Call = dyn_cast<CallInst>(it);
4213 if (Call && getIntrinsicIDForCall(Call, TLI))
4216 LoadInst *Ld = dyn_cast<LoadInst>(it);
4217 if (!Ld) return false;
4218 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4219 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4222 Loads.push_back(Ld);
4223 DepChecker.addAccess(Ld);
4227 // Save 'store' instructions. Abort if other instructions write to memory.
4228 if (it->mayWriteToMemory()) {
4229 StoreInst *St = dyn_cast<StoreInst>(it);
4230 if (!St) return false;
4231 if (!St->isSimple() && !IsAnnotatedParallel) {
4232 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4235 Stores.push_back(St);
4236 DepChecker.addAccess(St);
4241 // Now we have two lists that hold the loads and the stores.
4242 // Next, we find the pointers that they use.
4244 // Check if we see any stores. If there are no stores, then we don't
4245 // care if the pointers are *restrict*.
4246 if (!Stores.size()) {
4247 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4251 AccessAnalysis::DepCandidates DependentAccesses;
4252 AccessAnalysis Accesses(DL, DependentAccesses);
4254 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4255 // multiple times on the same object. If the ptr is accessed twice, once
4256 // for read and once for write, it will only appear once (on the write
4257 // list). This is okay, since we are going to check for conflicts between
4258 // writes and between reads and writes, but not between reads and reads.
4261 ValueVector::iterator I, IE;
4262 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4263 StoreInst *ST = cast<StoreInst>(*I);
4264 Value* Ptr = ST->getPointerOperand();
4266 if (isUniform(Ptr)) {
4267 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4271 // If we did *not* see this pointer before, insert it to the read-write
4272 // list. At this phase it is only a 'write' list.
4273 if (Seen.insert(Ptr)) {
4275 Accesses.addStore(Ptr);
4279 if (IsAnnotatedParallel) {
4281 << "LV: A loop annotated parallel, ignore memory dependency "
4286 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4287 LoadInst *LD = cast<LoadInst>(*I);
4288 Value* Ptr = LD->getPointerOperand();
4289 // If we did *not* see this pointer before, insert it to the
4290 // read list. If we *did* see it before, then it is already in
4291 // the read-write list. This allows us to vectorize expressions
4292 // such as A[i] += x; Because the address of A[i] is a read-write
4293 // pointer. This only works if the index of A[i] is consecutive.
4294 // If the address of i is unknown (for example A[B[i]]) then we may
4295 // read a few words, modify, and write a few words, and some of the
4296 // words may be written to the same address.
4297 bool IsReadOnlyPtr = false;
4298 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4300 IsReadOnlyPtr = true;
4302 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4305 // If we write (or read-write) to a single destination and there are no
4306 // other reads in this loop then is it safe to vectorize.
4307 if (NumReadWrites == 1 && NumReads == 0) {
4308 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4312 // Build dependence sets and check whether we need a runtime pointer bounds
4314 Accesses.buildDependenceSets();
4315 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4317 // Find pointers with computable bounds. We are going to use this information
4318 // to place a runtime bound check.
4319 unsigned NumComparisons = 0;
4320 bool CanDoRT = false;
4322 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4325 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4326 " pointer comparisons.\n");
4328 // If we only have one set of dependences to check pointers among we don't
4329 // need a runtime check.
4330 if (NumComparisons == 0 && NeedRTCheck)
4331 NeedRTCheck = false;
4333 // Check that we did not collect too many pointers or found an unsizeable
4335 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4341 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4344 if (NeedRTCheck && !CanDoRT) {
4345 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4346 "the array bounds.\n");
4351 PtrRtCheck.Need = NeedRTCheck;
4353 bool CanVecMem = true;
4354 if (Accesses.isDependencyCheckNeeded()) {
4355 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4356 CanVecMem = DepChecker.areDepsSafe(
4357 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4358 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4360 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4361 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4364 // Clear the dependency checks. We assume they are not needed.
4365 Accesses.resetDepChecks();
4368 PtrRtCheck.Need = true;
4370 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4371 TheLoop, Strides, true);
4372 // Check that we did not collect too many pointers or found an unsizeable
4374 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4375 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4384 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4385 " need a runtime memory check.\n");
4390 static bool hasMultipleUsesOf(Instruction *I,
4391 SmallPtrSet<Instruction *, 8> &Insts) {
4392 unsigned NumUses = 0;
4393 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4394 if (Insts.count(dyn_cast<Instruction>(*Use)))
4403 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4404 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4405 if (!Set.count(dyn_cast<Instruction>(*Use)))
4410 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4411 ReductionKind Kind) {
4412 if (Phi->getNumIncomingValues() != 2)
4415 // Reduction variables are only found in the loop header block.
4416 if (Phi->getParent() != TheLoop->getHeader())
4419 // Obtain the reduction start value from the value that comes from the loop
4421 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4423 // ExitInstruction is the single value which is used outside the loop.
4424 // We only allow for a single reduction value to be used outside the loop.
4425 // This includes users of the reduction, variables (which form a cycle
4426 // which ends in the phi node).
4427 Instruction *ExitInstruction = 0;
4428 // Indicates that we found a reduction operation in our scan.
4429 bool FoundReduxOp = false;
4431 // We start with the PHI node and scan for all of the users of this
4432 // instruction. All users must be instructions that can be used as reduction
4433 // variables (such as ADD). We must have a single out-of-block user. The cycle
4434 // must include the original PHI.
4435 bool FoundStartPHI = false;
4437 // To recognize min/max patterns formed by a icmp select sequence, we store
4438 // the number of instruction we saw from the recognized min/max pattern,
4439 // to make sure we only see exactly the two instructions.
4440 unsigned NumCmpSelectPatternInst = 0;
4441 ReductionInstDesc ReduxDesc(false, 0);
4443 SmallPtrSet<Instruction *, 8> VisitedInsts;
4444 SmallVector<Instruction *, 8> Worklist;
4445 Worklist.push_back(Phi);
4446 VisitedInsts.insert(Phi);
4448 // A value in the reduction can be used:
4449 // - By the reduction:
4450 // - Reduction operation:
4451 // - One use of reduction value (safe).
4452 // - Multiple use of reduction value (not safe).
4454 // - All uses of the PHI must be the reduction (safe).
4455 // - Otherwise, not safe.
4456 // - By one instruction outside of the loop (safe).
4457 // - By further instructions outside of the loop (not safe).
4458 // - By an instruction that is not part of the reduction (not safe).
4460 // * An instruction type other than PHI or the reduction operation.
4461 // * A PHI in the header other than the initial PHI.
4462 while (!Worklist.empty()) {
4463 Instruction *Cur = Worklist.back();
4464 Worklist.pop_back();
4467 // If the instruction has no users then this is a broken chain and can't be
4468 // a reduction variable.
4469 if (Cur->use_empty())
4472 bool IsAPhi = isa<PHINode>(Cur);
4474 // A header PHI use other than the original PHI.
4475 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4478 // Reductions of instructions such as Div, and Sub is only possible if the
4479 // LHS is the reduction variable.
4480 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4481 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4482 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4485 // Any reduction instruction must be of one of the allowed kinds.
4486 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4487 if (!ReduxDesc.IsReduction)
4490 // A reduction operation must only have one use of the reduction value.
4491 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4492 hasMultipleUsesOf(Cur, VisitedInsts))
4495 // All inputs to a PHI node must be a reduction value.
4496 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4499 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4500 isa<SelectInst>(Cur)))
4501 ++NumCmpSelectPatternInst;
4502 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4503 isa<SelectInst>(Cur)))
4504 ++NumCmpSelectPatternInst;
4506 // Check whether we found a reduction operator.
4507 FoundReduxOp |= !IsAPhi;
4509 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4510 // onto the stack. This way we are going to have seen all inputs to PHI
4511 // nodes once we get to them.
4512 SmallVector<Instruction *, 8> NonPHIs;
4513 SmallVector<Instruction *, 8> PHIs;
4514 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4516 Instruction *Usr = cast<Instruction>(*UI);
4518 // Check if we found the exit user.
4519 BasicBlock *Parent = Usr->getParent();
4520 if (!TheLoop->contains(Parent)) {
4521 // Exit if you find multiple outside users or if the header phi node is
4522 // being used. In this case the user uses the value of the previous
4523 // iteration, in which case we would loose "VF-1" iterations of the
4524 // reduction operation if we vectorize.
4525 if (ExitInstruction != 0 || Cur == Phi)
4528 // The instruction used by an outside user must be the last instruction
4529 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4530 // operations on the value.
4531 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4534 ExitInstruction = Cur;
4538 // Process instructions only once (termination). Each reduction cycle
4539 // value must only be used once, except by phi nodes and min/max
4540 // reductions which are represented as a cmp followed by a select.
4541 ReductionInstDesc IgnoredVal(false, 0);
4542 if (VisitedInsts.insert(Usr)) {
4543 if (isa<PHINode>(Usr))
4544 PHIs.push_back(Usr);
4546 NonPHIs.push_back(Usr);
4547 } else if (!isa<PHINode>(Usr) &&
4548 ((!isa<FCmpInst>(Usr) &&
4549 !isa<ICmpInst>(Usr) &&
4550 !isa<SelectInst>(Usr)) ||
4551 !isMinMaxSelectCmpPattern(Usr, IgnoredVal).IsReduction))
4554 // Remember that we completed the cycle.
4556 FoundStartPHI = true;
4558 Worklist.append(PHIs.begin(), PHIs.end());
4559 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4562 // This means we have seen one but not the other instruction of the
4563 // pattern or more than just a select and cmp.
4564 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4565 NumCmpSelectPatternInst != 2)
4568 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4571 // We found a reduction var if we have reached the original phi node and we
4572 // only have a single instruction with out-of-loop users.
4574 // This instruction is allowed to have out-of-loop users.
4575 AllowedExit.insert(ExitInstruction);
4577 // Save the description of this reduction variable.
4578 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4579 ReduxDesc.MinMaxKind);
4580 Reductions[Phi] = RD;
4581 // We've ended the cycle. This is a reduction variable if we have an
4582 // outside user and it has a binary op.
4587 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4588 /// pattern corresponding to a min(X, Y) or max(X, Y).
4589 LoopVectorizationLegality::ReductionInstDesc
4590 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4591 ReductionInstDesc &Prev) {
4593 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4594 "Expect a select instruction");
4595 Instruction *Cmp = 0;
4596 SelectInst *Select = 0;
4598 // We must handle the select(cmp()) as a single instruction. Advance to the
4600 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4601 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4602 return ReductionInstDesc(false, I);
4603 return ReductionInstDesc(Select, Prev.MinMaxKind);
4606 // Only handle single use cases for now.
4607 if (!(Select = dyn_cast<SelectInst>(I)))
4608 return ReductionInstDesc(false, I);
4609 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4610 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4611 return ReductionInstDesc(false, I);
4612 if (!Cmp->hasOneUse())
4613 return ReductionInstDesc(false, I);
4618 // Look for a min/max pattern.
4619 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4620 return ReductionInstDesc(Select, MRK_UIntMin);
4621 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4622 return ReductionInstDesc(Select, MRK_UIntMax);
4623 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4624 return ReductionInstDesc(Select, MRK_SIntMax);
4625 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4626 return ReductionInstDesc(Select, MRK_SIntMin);
4627 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4628 return ReductionInstDesc(Select, MRK_FloatMin);
4629 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4630 return ReductionInstDesc(Select, MRK_FloatMax);
4631 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4632 return ReductionInstDesc(Select, MRK_FloatMin);
4633 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4634 return ReductionInstDesc(Select, MRK_FloatMax);
4636 return ReductionInstDesc(false, I);
4639 LoopVectorizationLegality::ReductionInstDesc
4640 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4642 ReductionInstDesc &Prev) {
4643 bool FP = I->getType()->isFloatingPointTy();
4644 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4645 switch (I->getOpcode()) {
4647 return ReductionInstDesc(false, I);
4648 case Instruction::PHI:
4649 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4650 Kind != RK_FloatMinMax))
4651 return ReductionInstDesc(false, I);
4652 return ReductionInstDesc(I, Prev.MinMaxKind);
4653 case Instruction::Sub:
4654 case Instruction::Add:
4655 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4656 case Instruction::Mul:
4657 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4658 case Instruction::And:
4659 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4660 case Instruction::Or:
4661 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4662 case Instruction::Xor:
4663 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4664 case Instruction::FMul:
4665 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4666 case Instruction::FAdd:
4667 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4668 case Instruction::FCmp:
4669 case Instruction::ICmp:
4670 case Instruction::Select:
4671 if (Kind != RK_IntegerMinMax &&
4672 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4673 return ReductionInstDesc(false, I);
4674 return isMinMaxSelectCmpPattern(I, Prev);
4678 LoopVectorizationLegality::InductionKind
4679 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4680 Type *PhiTy = Phi->getType();
4681 // We only handle integer and pointer inductions variables.
4682 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4683 return IK_NoInduction;
4685 // Check that the PHI is consecutive.
4686 const SCEV *PhiScev = SE->getSCEV(Phi);
4687 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4689 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4690 return IK_NoInduction;
4692 const SCEV *Step = AR->getStepRecurrence(*SE);
4694 // Integer inductions need to have a stride of one.
4695 if (PhiTy->isIntegerTy()) {
4697 return IK_IntInduction;
4698 if (Step->isAllOnesValue())
4699 return IK_ReverseIntInduction;
4700 return IK_NoInduction;
4703 // Calculate the pointer stride and check if it is consecutive.
4704 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4706 return IK_NoInduction;
4708 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4709 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4710 if (C->getValue()->equalsInt(Size))
4711 return IK_PtrInduction;
4712 else if (C->getValue()->equalsInt(0 - Size))
4713 return IK_ReversePtrInduction;
4715 return IK_NoInduction;
4718 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4719 Value *In0 = const_cast<Value*>(V);
4720 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4724 return Inductions.count(PN);
4727 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4728 assert(TheLoop->contains(BB) && "Unknown block used");
4730 // Blocks that do not dominate the latch need predication.
4731 BasicBlock* Latch = TheLoop->getLoopLatch();
4732 return !DT->dominates(BB, Latch);
4735 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4736 SmallPtrSet<Value *, 8>& SafePtrs) {
4737 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4738 // We might be able to hoist the load.
4739 if (it->mayReadFromMemory()) {
4740 LoadInst *LI = dyn_cast<LoadInst>(it);
4741 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4745 // We don't predicate stores at the moment.
4746 if (it->mayWriteToMemory() || it->mayThrow())
4749 // Check that we don't have a constant expression that can trap as operand.
4750 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4752 if (Constant *C = dyn_cast<Constant>(*OI))
4757 // The instructions below can trap.
4758 switch (it->getOpcode()) {
4760 case Instruction::UDiv:
4761 case Instruction::SDiv:
4762 case Instruction::URem:
4763 case Instruction::SRem:
4771 LoopVectorizationCostModel::VectorizationFactor
4772 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4774 // Width 1 means no vectorize
4775 VectorizationFactor Factor = { 1U, 0U };
4776 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4777 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4781 // Find the trip count.
4782 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4783 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4785 unsigned WidestType = getWidestType();
4786 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4787 unsigned MaxSafeDepDist = -1U;
4788 if (Legal->getMaxSafeDepDistBytes() != -1U)
4789 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4790 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4791 WidestRegister : MaxSafeDepDist);
4792 unsigned MaxVectorSize = WidestRegister / WidestType;
4793 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4794 DEBUG(dbgs() << "LV: The Widest register is: "
4795 << WidestRegister << " bits.\n");
4797 if (MaxVectorSize == 0) {
4798 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4802 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4803 " into one vector!");
4805 unsigned VF = MaxVectorSize;
4807 // If we optimize the program for size, avoid creating the tail loop.
4809 // If we are unable to calculate the trip count then don't try to vectorize.
4811 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4815 // Find the maximum SIMD width that can fit within the trip count.
4816 VF = TC % MaxVectorSize;
4821 // If the trip count that we found modulo the vectorization factor is not
4822 // zero then we require a tail.
4824 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4830 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4831 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4833 Factor.Width = UserVF;
4837 float Cost = expectedCost(1);
4839 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
4840 for (unsigned i=2; i <= VF; i*=2) {
4841 // Notice that the vector loop needs to be executed less times, so
4842 // we need to divide the cost of the vector loops by the width of
4843 // the vector elements.
4844 float VectorCost = expectedCost(i) / (float)i;
4845 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4846 (int)VectorCost << ".\n");
4847 if (VectorCost < Cost) {
4853 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4854 Factor.Width = Width;
4855 Factor.Cost = Width * Cost;
4859 unsigned LoopVectorizationCostModel::getWidestType() {
4860 unsigned MaxWidth = 8;
4863 for (Loop::block_iterator bb = TheLoop->block_begin(),
4864 be = TheLoop->block_end(); bb != be; ++bb) {
4865 BasicBlock *BB = *bb;
4867 // For each instruction in the loop.
4868 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4869 Type *T = it->getType();
4871 // Only examine Loads, Stores and PHINodes.
4872 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4875 // Examine PHI nodes that are reduction variables.
4876 if (PHINode *PN = dyn_cast<PHINode>(it))
4877 if (!Legal->getReductionVars()->count(PN))
4880 // Examine the stored values.
4881 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4882 T = ST->getValueOperand()->getType();
4884 // Ignore loaded pointer types and stored pointer types that are not
4885 // consecutive. However, we do want to take consecutive stores/loads of
4886 // pointer vectors into account.
4887 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4890 MaxWidth = std::max(MaxWidth,
4891 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4899 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4902 unsigned LoopCost) {
4904 // -- The unroll heuristics --
4905 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4906 // There are many micro-architectural considerations that we can't predict
4907 // at this level. For example frontend pressure (on decode or fetch) due to
4908 // code size, or the number and capabilities of the execution ports.
4910 // We use the following heuristics to select the unroll factor:
4911 // 1. If the code has reductions the we unroll in order to break the cross
4912 // iteration dependency.
4913 // 2. If the loop is really small then we unroll in order to reduce the loop
4915 // 3. We don't unroll if we think that we will spill registers to memory due
4916 // to the increased register pressure.
4918 // Use the user preference, unless 'auto' is selected.
4922 // When we optimize for size we don't unroll.
4926 // We used the distance for the unroll factor.
4927 if (Legal->getMaxSafeDepDistBytes() != -1U)
4930 // Do not unroll loops with a relatively small trip count.
4931 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4932 TheLoop->getLoopLatch());
4933 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4936 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4937 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4938 " vector registers\n");
4940 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4941 // We divide by these constants so assume that we have at least one
4942 // instruction that uses at least one register.
4943 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4944 R.NumInstructions = std::max(R.NumInstructions, 1U);
4946 // We calculate the unroll factor using the following formula.
4947 // Subtract the number of loop invariants from the number of available
4948 // registers. These registers are used by all of the unrolled instances.
4949 // Next, divide the remaining registers by the number of registers that is
4950 // required by the loop, in order to estimate how many parallel instances
4951 // fit without causing spills.
4952 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4954 // Clamp the unroll factor ranges to reasonable factors.
4955 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4957 // If we did not calculate the cost for VF (because the user selected the VF)
4958 // then we calculate the cost of VF here.
4960 LoopCost = expectedCost(VF);
4962 // Clamp the calculated UF to be between the 1 and the max unroll factor
4963 // that the target allows.
4964 if (UF > MaxUnrollSize)
4969 bool HasReductions = Legal->getReductionVars()->size();
4971 // Decide if we want to unroll if we decided that it is legal to vectorize
4972 // but not profitable.
4974 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4975 LoopCost > SmallLoopCost)
4981 if (HasReductions) {
4982 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4986 // We want to unroll tiny loops in order to reduce the loop overhead.
4987 // We assume that the cost overhead is 1 and we use the cost model
4988 // to estimate the cost of the loop and unroll until the cost of the
4989 // loop overhead is about 5% of the cost of the loop.
4990 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4991 if (LoopCost < SmallLoopCost) {
4992 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4993 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4994 return std::min(NewUF, UF);
4997 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5001 LoopVectorizationCostModel::RegisterUsage
5002 LoopVectorizationCostModel::calculateRegisterUsage() {
5003 // This function calculates the register usage by measuring the highest number
5004 // of values that are alive at a single location. Obviously, this is a very
5005 // rough estimation. We scan the loop in a topological order in order and
5006 // assign a number to each instruction. We use RPO to ensure that defs are
5007 // met before their users. We assume that each instruction that has in-loop
5008 // users starts an interval. We record every time that an in-loop value is
5009 // used, so we have a list of the first and last occurrences of each
5010 // instruction. Next, we transpose this data structure into a multi map that
5011 // holds the list of intervals that *end* at a specific location. This multi
5012 // map allows us to perform a linear search. We scan the instructions linearly
5013 // and record each time that a new interval starts, by placing it in a set.
5014 // If we find this value in the multi-map then we remove it from the set.
5015 // The max register usage is the maximum size of the set.
5016 // We also search for instructions that are defined outside the loop, but are
5017 // used inside the loop. We need this number separately from the max-interval
5018 // usage number because when we unroll, loop-invariant values do not take
5020 LoopBlocksDFS DFS(TheLoop);
5024 R.NumInstructions = 0;
5026 // Each 'key' in the map opens a new interval. The values
5027 // of the map are the index of the 'last seen' usage of the
5028 // instruction that is the key.
5029 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5030 // Maps instruction to its index.
5031 DenseMap<unsigned, Instruction*> IdxToInstr;
5032 // Marks the end of each interval.
5033 IntervalMap EndPoint;
5034 // Saves the list of instruction indices that are used in the loop.
5035 SmallSet<Instruction*, 8> Ends;
5036 // Saves the list of values that are used in the loop but are
5037 // defined outside the loop, such as arguments and constants.
5038 SmallPtrSet<Value*, 8> LoopInvariants;
5041 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5042 be = DFS.endRPO(); bb != be; ++bb) {
5043 R.NumInstructions += (*bb)->size();
5044 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5046 Instruction *I = it;
5047 IdxToInstr[Index++] = I;
5049 // Save the end location of each USE.
5050 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5051 Value *U = I->getOperand(i);
5052 Instruction *Instr = dyn_cast<Instruction>(U);
5054 // Ignore non-instruction values such as arguments, constants, etc.
5055 if (!Instr) continue;
5057 // If this instruction is outside the loop then record it and continue.
5058 if (!TheLoop->contains(Instr)) {
5059 LoopInvariants.insert(Instr);
5063 // Overwrite previous end points.
5064 EndPoint[Instr] = Index;
5070 // Saves the list of intervals that end with the index in 'key'.
5071 typedef SmallVector<Instruction*, 2> InstrList;
5072 DenseMap<unsigned, InstrList> TransposeEnds;
5074 // Transpose the EndPoints to a list of values that end at each index.
5075 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5077 TransposeEnds[it->second].push_back(it->first);
5079 SmallSet<Instruction*, 8> OpenIntervals;
5080 unsigned MaxUsage = 0;
5083 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5084 for (unsigned int i = 0; i < Index; ++i) {
5085 Instruction *I = IdxToInstr[i];
5086 // Ignore instructions that are never used within the loop.
5087 if (!Ends.count(I)) continue;
5089 // Remove all of the instructions that end at this location.
5090 InstrList &List = TransposeEnds[i];
5091 for (unsigned int j=0, e = List.size(); j < e; ++j)
5092 OpenIntervals.erase(List[j]);
5094 // Count the number of live interals.
5095 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5097 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5098 OpenIntervals.size() << '\n');
5100 // Add the current instruction to the list of open intervals.
5101 OpenIntervals.insert(I);
5104 unsigned Invariant = LoopInvariants.size();
5105 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5106 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5107 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5109 R.LoopInvariantRegs = Invariant;
5110 R.MaxLocalUsers = MaxUsage;
5114 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5118 for (Loop::block_iterator bb = TheLoop->block_begin(),
5119 be = TheLoop->block_end(); bb != be; ++bb) {
5120 unsigned BlockCost = 0;
5121 BasicBlock *BB = *bb;
5123 // For each instruction in the old loop.
5124 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5125 // Skip dbg intrinsics.
5126 if (isa<DbgInfoIntrinsic>(it))
5129 unsigned C = getInstructionCost(it, VF);
5131 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5132 VF << " For instruction: " << *it << '\n');
5135 // We assume that if-converted blocks have a 50% chance of being executed.
5136 // When the code is scalar then some of the blocks are avoided due to CF.
5137 // When the code is vectorized we execute all code paths.
5138 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5147 /// \brief Check whether the address computation for a non-consecutive memory
5148 /// access looks like an unlikely candidate for being merged into the indexing
5151 /// We look for a GEP which has one index that is an induction variable and all
5152 /// other indices are loop invariant. If the stride of this access is also
5153 /// within a small bound we decide that this address computation can likely be
5154 /// merged into the addressing mode.
5155 /// In all other cases, we identify the address computation as complex.
5156 static bool isLikelyComplexAddressComputation(Value *Ptr,
5157 LoopVectorizationLegality *Legal,
5158 ScalarEvolution *SE,
5159 const Loop *TheLoop) {
5160 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5164 // We are looking for a gep with all loop invariant indices except for one
5165 // which should be an induction variable.
5166 unsigned NumOperands = Gep->getNumOperands();
5167 for (unsigned i = 1; i < NumOperands; ++i) {
5168 Value *Opd = Gep->getOperand(i);
5169 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5170 !Legal->isInductionVariable(Opd))
5174 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5175 // can likely be merged into the address computation.
5176 unsigned MaxMergeDistance = 64;
5178 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5182 // Check the step is constant.
5183 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5184 // Calculate the pointer stride and check if it is consecutive.
5185 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5189 const APInt &APStepVal = C->getValue()->getValue();
5191 // Huge step value - give up.
5192 if (APStepVal.getBitWidth() > 64)
5195 int64_t StepVal = APStepVal.getSExtValue();
5197 return StepVal > MaxMergeDistance;
5200 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5201 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5207 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5208 // If we know that this instruction will remain uniform, check the cost of
5209 // the scalar version.
5210 if (Legal->isUniformAfterVectorization(I))
5213 Type *RetTy = I->getType();
5214 Type *VectorTy = ToVectorTy(RetTy, VF);
5216 // TODO: We need to estimate the cost of intrinsic calls.
5217 switch (I->getOpcode()) {
5218 case Instruction::GetElementPtr:
5219 // We mark this instruction as zero-cost because the cost of GEPs in
5220 // vectorized code depends on whether the corresponding memory instruction
5221 // is scalarized or not. Therefore, we handle GEPs with the memory
5222 // instruction cost.
5224 case Instruction::Br: {
5225 return TTI.getCFInstrCost(I->getOpcode());
5227 case Instruction::PHI:
5228 //TODO: IF-converted IFs become selects.
5230 case Instruction::Add:
5231 case Instruction::FAdd:
5232 case Instruction::Sub:
5233 case Instruction::FSub:
5234 case Instruction::Mul:
5235 case Instruction::FMul:
5236 case Instruction::UDiv:
5237 case Instruction::SDiv:
5238 case Instruction::FDiv:
5239 case Instruction::URem:
5240 case Instruction::SRem:
5241 case Instruction::FRem:
5242 case Instruction::Shl:
5243 case Instruction::LShr:
5244 case Instruction::AShr:
5245 case Instruction::And:
5246 case Instruction::Or:
5247 case Instruction::Xor: {
5248 // Since we will replace the stride by 1 the multiplication should go away.
5249 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5251 // Certain instructions can be cheaper to vectorize if they have a constant
5252 // second vector operand. One example of this are shifts on x86.
5253 TargetTransformInfo::OperandValueKind Op1VK =
5254 TargetTransformInfo::OK_AnyValue;
5255 TargetTransformInfo::OperandValueKind Op2VK =
5256 TargetTransformInfo::OK_AnyValue;
5258 if (isa<ConstantInt>(I->getOperand(1)))
5259 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5261 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5263 case Instruction::Select: {
5264 SelectInst *SI = cast<SelectInst>(I);
5265 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5266 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5267 Type *CondTy = SI->getCondition()->getType();
5269 CondTy = VectorType::get(CondTy, VF);
5271 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5273 case Instruction::ICmp:
5274 case Instruction::FCmp: {
5275 Type *ValTy = I->getOperand(0)->getType();
5276 VectorTy = ToVectorTy(ValTy, VF);
5277 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5279 case Instruction::Store:
5280 case Instruction::Load: {
5281 StoreInst *SI = dyn_cast<StoreInst>(I);
5282 LoadInst *LI = dyn_cast<LoadInst>(I);
5283 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5285 VectorTy = ToVectorTy(ValTy, VF);
5287 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5288 unsigned AS = SI ? SI->getPointerAddressSpace() :
5289 LI->getPointerAddressSpace();
5290 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5291 // We add the cost of address computation here instead of with the gep
5292 // instruction because only here we know whether the operation is
5295 return TTI.getAddressComputationCost(VectorTy) +
5296 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5298 // Scalarized loads/stores.
5299 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5300 bool Reverse = ConsecutiveStride < 0;
5301 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5302 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5303 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5304 bool IsComplexComputation =
5305 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5307 // The cost of extracting from the value vector and pointer vector.
5308 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5309 for (unsigned i = 0; i < VF; ++i) {
5310 // The cost of extracting the pointer operand.
5311 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5312 // In case of STORE, the cost of ExtractElement from the vector.
5313 // In case of LOAD, the cost of InsertElement into the returned
5315 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5316 Instruction::InsertElement,
5320 // The cost of the scalar loads/stores.
5321 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5322 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5327 // Wide load/stores.
5328 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5329 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5332 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5336 case Instruction::ZExt:
5337 case Instruction::SExt:
5338 case Instruction::FPToUI:
5339 case Instruction::FPToSI:
5340 case Instruction::FPExt:
5341 case Instruction::PtrToInt:
5342 case Instruction::IntToPtr:
5343 case Instruction::SIToFP:
5344 case Instruction::UIToFP:
5345 case Instruction::Trunc:
5346 case Instruction::FPTrunc:
5347 case Instruction::BitCast: {
5348 // We optimize the truncation of induction variable.
5349 // The cost of these is the same as the scalar operation.
5350 if (I->getOpcode() == Instruction::Trunc &&
5351 Legal->isInductionVariable(I->getOperand(0)))
5352 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5353 I->getOperand(0)->getType());
5355 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5356 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5358 case Instruction::Call: {
5359 CallInst *CI = cast<CallInst>(I);
5360 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5361 assert(ID && "Not an intrinsic call!");
5362 Type *RetTy = ToVectorTy(CI->getType(), VF);
5363 SmallVector<Type*, 4> Tys;
5364 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5365 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5366 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5369 // We are scalarizing the instruction. Return the cost of the scalar
5370 // instruction, plus the cost of insert and extract into vector
5371 // elements, times the vector width.
5374 if (!RetTy->isVoidTy() && VF != 1) {
5375 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5377 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5380 // The cost of inserting the results plus extracting each one of the
5382 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5385 // The cost of executing VF copies of the scalar instruction. This opcode
5386 // is unknown. Assume that it is the same as 'mul'.
5387 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5393 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5394 if (Scalar->isVoidTy() || VF == 1)
5396 return VectorType::get(Scalar, VF);
5399 char LoopVectorize::ID = 0;
5400 static const char lv_name[] = "Loop Vectorization";
5401 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5402 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5403 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5404 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5405 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5406 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5407 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5408 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5411 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5412 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5416 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5417 // Check for a store.
5418 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5419 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5421 // Check for a load.
5422 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5423 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5429 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
5430 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5431 // Holds vector parameters or scalars, in case of uniform vals.
5432 SmallVector<VectorParts, 4> Params;
5434 setDebugLocFromInst(Builder, Instr);
5436 // Find all of the vectorized parameters.
5437 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5438 Value *SrcOp = Instr->getOperand(op);
5440 // If we are accessing the old induction variable, use the new one.
5441 if (SrcOp == OldInduction) {
5442 Params.push_back(getVectorValue(SrcOp));
5446 // Try using previously calculated values.
5447 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5449 // If the src is an instruction that appeared earlier in the basic block
5450 // then it should already be vectorized.
5451 if (SrcInst && OrigLoop->contains(SrcInst)) {
5452 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5453 // The parameter is a vector value from earlier.
5454 Params.push_back(WidenMap.get(SrcInst));
5456 // The parameter is a scalar from outside the loop. Maybe even a constant.
5457 VectorParts Scalars;
5458 Scalars.append(UF, SrcOp);
5459 Params.push_back(Scalars);
5463 assert(Params.size() == Instr->getNumOperands() &&
5464 "Invalid number of operands");
5466 // Does this instruction return a value ?
5467 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5469 Value *UndefVec = IsVoidRetTy ? 0 :
5470 UndefValue::get(Instr->getType());
5471 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5472 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5474 // For each vector unroll 'part':
5475 for (unsigned Part = 0; Part < UF; ++Part) {
5476 // For each scalar that we create:
5478 Instruction *Cloned = Instr->clone();
5480 Cloned->setName(Instr->getName() + ".cloned");
5481 // Replace the operands of the cloned instructions with extracted scalars.
5482 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5483 Value *Op = Params[op][Part];
5484 Cloned->setOperand(op, Op);
5487 // Place the cloned scalar in the new loop.
5488 Builder.Insert(Cloned);
5490 // If the original scalar returns a value we need to place it in a vector
5491 // so that future users will be able to use it.
5493 VecResults[Part] = Cloned;
5497 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5498 return scalarizeInstruction(Instr);
5501 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5505 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5509 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5511 // When unrolling and the VF is 1, we only need to add a simple scalar.
5512 Type *ITy = Val->getType();
5513 assert(!ITy->isVectorTy() && "Val must be a scalar");
5514 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5515 return Builder.CreateAdd(Val, C, "induction");