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/Dominators.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/Analysis/Verifier.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DerivedTypes.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.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(false), 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<DominatorTree>();
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<DominatorTree>();
1059 AU.addRequired<LoopInfo>();
1060 AU.addRequired<ScalarEvolution>();
1061 AU.addRequired<TargetTransformInfo>();
1062 AU.addPreserved<LoopInfo>();
1063 AU.addPreserved<DominatorTree>();
1068 } // end anonymous namespace
1070 //===----------------------------------------------------------------------===//
1071 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1072 // LoopVectorizationCostModel.
1073 //===----------------------------------------------------------------------===//
1075 static Value *stripCast(Value *V) {
1076 if (CastInst *CI = dyn_cast<CastInst>(V))
1077 return CI->getOperand(0);
1081 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1083 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1085 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1086 ValueToValueMap &PtrToStride,
1087 Value *Ptr, Value *OrigPtr = 0) {
1089 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1091 // If there is an entry in the map return the SCEV of the pointer with the
1092 // symbolic stride replaced by one.
1093 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1094 if (SI != PtrToStride.end()) {
1095 Value *StrideVal = SI->second;
1098 StrideVal = stripCast(StrideVal);
1100 // Replace symbolic stride by one.
1101 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1102 ValueToValueMap RewriteMap;
1103 RewriteMap[StrideVal] = One;
1106 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1107 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1112 // Otherwise, just return the SCEV of the original pointer.
1113 return SE->getSCEV(Ptr);
1116 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1117 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1118 ValueToValueMap &Strides) {
1119 // Get the stride replaced scev.
1120 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1121 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1122 assert(AR && "Invalid addrec expression");
1123 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1124 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1125 Pointers.push_back(Ptr);
1126 Starts.push_back(AR->getStart());
1127 Ends.push_back(ScEnd);
1128 IsWritePtr.push_back(WritePtr);
1129 DependencySetId.push_back(DepSetId);
1132 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1133 // We need to place the broadcast of invariant variables outside the loop.
1134 Instruction *Instr = dyn_cast<Instruction>(V);
1135 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1136 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1138 // Place the code for broadcasting invariant variables in the new preheader.
1139 IRBuilder<>::InsertPointGuard Guard(Builder);
1141 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1143 // Broadcast the scalar into all locations in the vector.
1144 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1149 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1151 assert(Val->getType()->isVectorTy() && "Must be a vector");
1152 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1153 "Elem must be an integer");
1154 // Create the types.
1155 Type *ITy = Val->getType()->getScalarType();
1156 VectorType *Ty = cast<VectorType>(Val->getType());
1157 int VLen = Ty->getNumElements();
1158 SmallVector<Constant*, 8> Indices;
1160 // Create a vector of consecutive numbers from zero to VF.
1161 for (int i = 0; i < VLen; ++i) {
1162 int64_t Idx = Negate ? (-i) : i;
1163 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1166 // Add the consecutive indices to the vector value.
1167 Constant *Cv = ConstantVector::get(Indices);
1168 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1169 return Builder.CreateAdd(Val, Cv, "induction");
1172 /// \brief Find the operand of the GEP that should be checked for consecutive
1173 /// stores. This ignores trailing indices that have no effect on the final
1175 static unsigned getGEPInductionOperand(DataLayout *DL,
1176 const GetElementPtrInst *Gep) {
1177 unsigned LastOperand = Gep->getNumOperands() - 1;
1178 unsigned GEPAllocSize = DL->getTypeAllocSize(
1179 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1181 // Walk backwards and try to peel off zeros.
1182 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1183 // Find the type we're currently indexing into.
1184 gep_type_iterator GEPTI = gep_type_begin(Gep);
1185 std::advance(GEPTI, LastOperand - 1);
1187 // If it's a type with the same allocation size as the result of the GEP we
1188 // can peel off the zero index.
1189 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1197 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1198 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1199 // Make sure that the pointer does not point to structs.
1200 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1203 // If this value is a pointer induction variable we know it is consecutive.
1204 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1205 if (Phi && Inductions.count(Phi)) {
1206 InductionInfo II = Inductions[Phi];
1207 if (IK_PtrInduction == II.IK)
1209 else if (IK_ReversePtrInduction == II.IK)
1213 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1217 unsigned NumOperands = Gep->getNumOperands();
1218 Value *GpPtr = Gep->getPointerOperand();
1219 // If this GEP value is a consecutive pointer induction variable and all of
1220 // the indices are constant then we know it is consecutive. We can
1221 Phi = dyn_cast<PHINode>(GpPtr);
1222 if (Phi && Inductions.count(Phi)) {
1224 // Make sure that the pointer does not point to structs.
1225 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1226 if (GepPtrType->getElementType()->isAggregateType())
1229 // Make sure that all of the index operands are loop invariant.
1230 for (unsigned i = 1; i < NumOperands; ++i)
1231 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1234 InductionInfo II = Inductions[Phi];
1235 if (IK_PtrInduction == II.IK)
1237 else if (IK_ReversePtrInduction == II.IK)
1241 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1243 // Check that all of the gep indices are uniform except for our induction
1245 for (unsigned i = 0; i != NumOperands; ++i)
1246 if (i != InductionOperand &&
1247 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1250 // We can emit wide load/stores only if the last non-zero index is the
1251 // induction variable.
1252 const SCEV *Last = 0;
1253 if (!Strides.count(Gep))
1254 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1256 // Because of the multiplication by a stride we can have a s/zext cast.
1257 // We are going to replace this stride by 1 so the cast is safe to ignore.
1259 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1260 // %0 = trunc i64 %indvars.iv to i32
1261 // %mul = mul i32 %0, %Stride1
1262 // %idxprom = zext i32 %mul to i64 << Safe cast.
1263 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1265 Last = replaceSymbolicStrideSCEV(SE, Strides,
1266 Gep->getOperand(InductionOperand), Gep);
1267 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1269 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1273 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1274 const SCEV *Step = AR->getStepRecurrence(*SE);
1276 // The memory is consecutive because the last index is consecutive
1277 // and all other indices are loop invariant.
1280 if (Step->isAllOnesValue())
1287 bool LoopVectorizationLegality::isUniform(Value *V) {
1288 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1291 InnerLoopVectorizer::VectorParts&
1292 InnerLoopVectorizer::getVectorValue(Value *V) {
1293 assert(V != Induction && "The new induction variable should not be used.");
1294 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1296 // If we have a stride that is replaced by one, do it here.
1297 if (Legal->hasStride(V))
1298 V = ConstantInt::get(V->getType(), 1);
1300 // If we have this scalar in the map, return it.
1301 if (WidenMap.has(V))
1302 return WidenMap.get(V);
1304 // If this scalar is unknown, assume that it is a constant or that it is
1305 // loop invariant. Broadcast V and save the value for future uses.
1306 Value *B = getBroadcastInstrs(V);
1307 return WidenMap.splat(V, B);
1310 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1311 assert(Vec->getType()->isVectorTy() && "Invalid type");
1312 SmallVector<Constant*, 8> ShuffleMask;
1313 for (unsigned i = 0; i < VF; ++i)
1314 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1316 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1317 ConstantVector::get(ShuffleMask),
1321 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1322 // Attempt to issue a wide load.
1323 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1324 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1326 assert((LI || SI) && "Invalid Load/Store instruction");
1328 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1329 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1330 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1331 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1332 // An alignment of 0 means target abi alignment. We need to use the scalar's
1333 // target abi alignment in such a case.
1335 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1336 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1337 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1338 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1340 if (ScalarAllocatedSize != VectorElementSize)
1341 return scalarizeInstruction(Instr);
1343 // If the pointer is loop invariant or if it is non-consecutive,
1344 // scalarize the load.
1345 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1346 bool Reverse = ConsecutiveStride < 0;
1347 bool UniformLoad = LI && Legal->isUniform(Ptr);
1348 if (!ConsecutiveStride || UniformLoad)
1349 return scalarizeInstruction(Instr);
1351 Constant *Zero = Builder.getInt32(0);
1352 VectorParts &Entry = WidenMap.get(Instr);
1354 // Handle consecutive loads/stores.
1355 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1356 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1357 setDebugLocFromInst(Builder, Gep);
1358 Value *PtrOperand = Gep->getPointerOperand();
1359 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1360 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1362 // Create the new GEP with the new induction variable.
1363 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1364 Gep2->setOperand(0, FirstBasePtr);
1365 Gep2->setName("gep.indvar.base");
1366 Ptr = Builder.Insert(Gep2);
1368 setDebugLocFromInst(Builder, Gep);
1369 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1370 OrigLoop) && "Base ptr must be invariant");
1372 // The last index does not have to be the induction. It can be
1373 // consecutive and be a function of the index. For example A[I+1];
1374 unsigned NumOperands = Gep->getNumOperands();
1375 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1376 // Create the new GEP with the new induction variable.
1377 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1379 for (unsigned i = 0; i < NumOperands; ++i) {
1380 Value *GepOperand = Gep->getOperand(i);
1381 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1383 // Update last index or loop invariant instruction anchored in loop.
1384 if (i == InductionOperand ||
1385 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1386 assert((i == InductionOperand ||
1387 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1388 "Must be last index or loop invariant");
1390 VectorParts &GEPParts = getVectorValue(GepOperand);
1391 Value *Index = GEPParts[0];
1392 Index = Builder.CreateExtractElement(Index, Zero);
1393 Gep2->setOperand(i, Index);
1394 Gep2->setName("gep.indvar.idx");
1397 Ptr = Builder.Insert(Gep2);
1399 // Use the induction element ptr.
1400 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1401 setDebugLocFromInst(Builder, Ptr);
1402 VectorParts &PtrVal = getVectorValue(Ptr);
1403 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1408 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1409 "We do not allow storing to uniform addresses");
1410 setDebugLocFromInst(Builder, SI);
1411 // We don't want to update the value in the map as it might be used in
1412 // another expression. So don't use a reference type for "StoredVal".
1413 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1415 for (unsigned Part = 0; Part < UF; ++Part) {
1416 // Calculate the pointer for the specific unroll-part.
1417 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1420 // If we store to reverse consecutive memory locations then we need
1421 // to reverse the order of elements in the stored value.
1422 StoredVal[Part] = reverseVector(StoredVal[Part]);
1423 // If the address is consecutive but reversed, then the
1424 // wide store needs to start at the last vector element.
1425 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1426 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1429 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1430 DataTy->getPointerTo(AddressSpace));
1431 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1437 assert(LI && "Must have a load instruction");
1438 setDebugLocFromInst(Builder, LI);
1439 for (unsigned Part = 0; Part < UF; ++Part) {
1440 // Calculate the pointer for the specific unroll-part.
1441 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1444 // If the address is consecutive but reversed, then the
1445 // wide store needs to start at the last vector element.
1446 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1447 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1450 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1451 DataTy->getPointerTo(AddressSpace));
1452 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1453 cast<LoadInst>(LI)->setAlignment(Alignment);
1454 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1458 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1459 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1460 // Holds vector parameters or scalars, in case of uniform vals.
1461 SmallVector<VectorParts, 4> Params;
1463 setDebugLocFromInst(Builder, Instr);
1465 // Find all of the vectorized parameters.
1466 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1467 Value *SrcOp = Instr->getOperand(op);
1469 // If we are accessing the old induction variable, use the new one.
1470 if (SrcOp == OldInduction) {
1471 Params.push_back(getVectorValue(SrcOp));
1475 // Try using previously calculated values.
1476 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1478 // If the src is an instruction that appeared earlier in the basic block
1479 // then it should already be vectorized.
1480 if (SrcInst && OrigLoop->contains(SrcInst)) {
1481 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1482 // The parameter is a vector value from earlier.
1483 Params.push_back(WidenMap.get(SrcInst));
1485 // The parameter is a scalar from outside the loop. Maybe even a constant.
1486 VectorParts Scalars;
1487 Scalars.append(UF, SrcOp);
1488 Params.push_back(Scalars);
1492 assert(Params.size() == Instr->getNumOperands() &&
1493 "Invalid number of operands");
1495 // Does this instruction return a value ?
1496 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1498 Value *UndefVec = IsVoidRetTy ? 0 :
1499 UndefValue::get(VectorType::get(Instr->getType(), VF));
1500 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1501 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1503 // For each vector unroll 'part':
1504 for (unsigned Part = 0; Part < UF; ++Part) {
1505 // For each scalar that we create:
1506 for (unsigned Width = 0; Width < VF; ++Width) {
1507 Instruction *Cloned = Instr->clone();
1509 Cloned->setName(Instr->getName() + ".cloned");
1510 // Replace the operands of the cloned instructions with extracted scalars.
1511 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1512 Value *Op = Params[op][Part];
1513 // Param is a vector. Need to extract the right lane.
1514 if (Op->getType()->isVectorTy())
1515 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1516 Cloned->setOperand(op, Op);
1519 // Place the cloned scalar in the new loop.
1520 Builder.Insert(Cloned);
1522 // If the original scalar returns a value we need to place it in a vector
1523 // so that future users will be able to use it.
1525 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1526 Builder.getInt32(Width));
1531 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1535 if (Instruction *I = dyn_cast<Instruction>(V))
1536 return I->getParent() == Loc->getParent() ? I : 0;
1540 std::pair<Instruction *, Instruction *>
1541 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1542 if (!Legal->mustCheckStrides())
1543 return std::pair<Instruction *, Instruction *>(0, 0);
1545 IRBuilder<> ChkBuilder(Loc);
1549 Instruction *FirstInst = 0;
1550 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1551 SE = Legal->strides_end();
1553 Value *Ptr = stripCast(*SI);
1554 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1556 // Store the first instruction we create.
1557 FirstInst = getFirstInst(FirstInst, C, Loc);
1559 Check = ChkBuilder.CreateOr(Check, C);
1564 // We have to do this trickery because the IRBuilder might fold the check to a
1565 // constant expression in which case there is no Instruction anchored in a
1567 LLVMContext &Ctx = Loc->getContext();
1568 Instruction *TheCheck =
1569 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1570 ChkBuilder.Insert(TheCheck, "stride.not.one");
1571 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1573 return std::make_pair(FirstInst, TheCheck);
1576 std::pair<Instruction *, Instruction *>
1577 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1578 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1579 Legal->getRuntimePointerCheck();
1581 if (!PtrRtCheck->Need)
1582 return std::pair<Instruction *, Instruction *>(0, 0);
1584 unsigned NumPointers = PtrRtCheck->Pointers.size();
1585 SmallVector<TrackingVH<Value> , 2> Starts;
1586 SmallVector<TrackingVH<Value> , 2> Ends;
1588 LLVMContext &Ctx = Loc->getContext();
1589 SCEVExpander Exp(*SE, "induction");
1590 Instruction *FirstInst = 0;
1592 for (unsigned i = 0; i < NumPointers; ++i) {
1593 Value *Ptr = PtrRtCheck->Pointers[i];
1594 const SCEV *Sc = SE->getSCEV(Ptr);
1596 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1597 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1599 Starts.push_back(Ptr);
1600 Ends.push_back(Ptr);
1602 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1603 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1605 // Use this type for pointer arithmetic.
1606 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1608 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1609 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1610 Starts.push_back(Start);
1611 Ends.push_back(End);
1615 IRBuilder<> ChkBuilder(Loc);
1616 // Our instructions might fold to a constant.
1617 Value *MemoryRuntimeCheck = 0;
1618 for (unsigned i = 0; i < NumPointers; ++i) {
1619 for (unsigned j = i+1; j < NumPointers; ++j) {
1620 // No need to check if two readonly pointers intersect.
1621 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1624 // Only need to check pointers between two different dependency sets.
1625 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1628 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1629 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1631 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1632 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1633 "Trying to bounds check pointers with different address spaces");
1635 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1636 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1638 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1639 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1640 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1641 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1643 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1644 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1645 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1646 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1647 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1648 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1649 if (MemoryRuntimeCheck) {
1650 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1652 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1654 MemoryRuntimeCheck = IsConflict;
1658 // We have to do this trickery because the IRBuilder might fold the check to a
1659 // constant expression in which case there is no Instruction anchored in a
1661 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1662 ConstantInt::getTrue(Ctx));
1663 ChkBuilder.Insert(Check, "memcheck.conflict");
1664 FirstInst = getFirstInst(FirstInst, Check, Loc);
1665 return std::make_pair(FirstInst, Check);
1668 void InnerLoopVectorizer::createEmptyLoop() {
1670 In this function we generate a new loop. The new loop will contain
1671 the vectorized instructions while the old loop will continue to run the
1674 [ ] <-- vector loop bypass (may consist of multiple blocks).
1677 | [ ] <-- vector pre header.
1681 | [ ]_| <-- vector loop.
1684 >[ ] <--- middle-block.
1687 | [ ] <--- new preheader.
1691 | [ ]_| <-- old scalar loop to handle remainder.
1694 >[ ] <-- exit block.
1698 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1699 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1700 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1701 assert(ExitBlock && "Must have an exit block");
1703 // Some loops have a single integer induction variable, while other loops
1704 // don't. One example is c++ iterators that often have multiple pointer
1705 // induction variables. In the code below we also support a case where we
1706 // don't have a single induction variable.
1707 OldInduction = Legal->getInduction();
1708 Type *IdxTy = Legal->getWidestInductionType();
1710 // Find the loop boundaries.
1711 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1712 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1714 // The exit count might have the type of i64 while the phi is i32. This can
1715 // happen if we have an induction variable that is sign extended before the
1716 // compare. The only way that we get a backedge taken count is that the
1717 // induction variable was signed and as such will not overflow. In such a case
1718 // truncation is legal.
1719 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1720 IdxTy->getPrimitiveSizeInBits())
1721 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1723 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1724 // Get the total trip count from the count by adding 1.
1725 ExitCount = SE->getAddExpr(ExitCount,
1726 SE->getConstant(ExitCount->getType(), 1));
1728 // Expand the trip count and place the new instructions in the preheader.
1729 // Notice that the pre-header does not change, only the loop body.
1730 SCEVExpander Exp(*SE, "induction");
1732 // Count holds the overall loop count (N).
1733 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1734 BypassBlock->getTerminator());
1736 // The loop index does not have to start at Zero. Find the original start
1737 // value from the induction PHI node. If we don't have an induction variable
1738 // then we know that it starts at zero.
1739 Builder.SetInsertPoint(BypassBlock->getTerminator());
1740 Value *StartIdx = ExtendedIdx = OldInduction ?
1741 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1743 ConstantInt::get(IdxTy, 0);
1745 assert(BypassBlock && "Invalid loop structure");
1746 LoopBypassBlocks.push_back(BypassBlock);
1748 // Split the single block loop into the two loop structure described above.
1749 BasicBlock *VectorPH =
1750 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1751 BasicBlock *VecBody =
1752 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1753 BasicBlock *MiddleBlock =
1754 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1755 BasicBlock *ScalarPH =
1756 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1758 // Create and register the new vector loop.
1759 Loop* Lp = new Loop();
1760 Loop *ParentLoop = OrigLoop->getParentLoop();
1762 // Insert the new loop into the loop nest and register the new basic blocks
1763 // before calling any utilities such as SCEV that require valid LoopInfo.
1765 ParentLoop->addChildLoop(Lp);
1766 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1767 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1768 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1770 LI->addTopLevelLoop(Lp);
1772 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1774 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1776 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1778 // Generate the induction variable.
1779 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1780 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1781 // The loop step is equal to the vectorization factor (num of SIMD elements)
1782 // times the unroll factor (num of SIMD instructions).
1783 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1785 // This is the IR builder that we use to add all of the logic for bypassing
1786 // the new vector loop.
1787 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1788 setDebugLocFromInst(BypassBuilder,
1789 getDebugLocFromInstOrOperands(OldInduction));
1791 // We may need to extend the index in case there is a type mismatch.
1792 // We know that the count starts at zero and does not overflow.
1793 if (Count->getType() != IdxTy) {
1794 // The exit count can be of pointer type. Convert it to the correct
1796 if (ExitCount->getType()->isPointerTy())
1797 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1799 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1802 // Add the start index to the loop count to get the new end index.
1803 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1805 // Now we need to generate the expression for N - (N % VF), which is
1806 // the part that the vectorized body will execute.
1807 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1808 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1809 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1810 "end.idx.rnd.down");
1812 // Now, compare the new count to zero. If it is zero skip the vector loop and
1813 // jump to the scalar loop.
1814 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1817 BasicBlock *LastBypassBlock = BypassBlock;
1819 // Generate the code to check that the strides we assumed to be one are really
1820 // one. We want the new basic block to start at the first instruction in a
1821 // sequence of instructions that form a check.
1822 Instruction *StrideCheck;
1823 Instruction *FirstCheckInst;
1824 tie(FirstCheckInst, StrideCheck) =
1825 addStrideCheck(BypassBlock->getTerminator());
1827 // Create a new block containing the stride check.
1828 BasicBlock *CheckBlock =
1829 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
1831 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1832 LoopBypassBlocks.push_back(CheckBlock);
1834 // Replace the branch into the memory check block with a conditional branch
1835 // for the "few elements case".
1836 Instruction *OldTerm = BypassBlock->getTerminator();
1837 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1838 OldTerm->eraseFromParent();
1841 LastBypassBlock = CheckBlock;
1844 // Generate the code that checks in runtime if arrays overlap. We put the
1845 // checks into a separate block to make the more common case of few elements
1847 Instruction *MemRuntimeCheck;
1848 tie(FirstCheckInst, MemRuntimeCheck) =
1849 addRuntimeCheck(LastBypassBlock->getTerminator());
1850 if (MemRuntimeCheck) {
1851 // Create a new block containing the memory check.
1852 BasicBlock *CheckBlock =
1853 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
1855 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1856 LoopBypassBlocks.push_back(CheckBlock);
1858 // Replace the branch into the memory check block with a conditional branch
1859 // for the "few elements case".
1860 Instruction *OldTerm = LastBypassBlock->getTerminator();
1861 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1862 OldTerm->eraseFromParent();
1864 Cmp = MemRuntimeCheck;
1865 LastBypassBlock = CheckBlock;
1868 LastBypassBlock->getTerminator()->eraseFromParent();
1869 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1872 // We are going to resume the execution of the scalar loop.
1873 // Go over all of the induction variables that we found and fix the
1874 // PHIs that are left in the scalar version of the loop.
1875 // The starting values of PHI nodes depend on the counter of the last
1876 // iteration in the vectorized loop.
1877 // If we come from a bypass edge then we need to start from the original
1880 // This variable saves the new starting index for the scalar loop.
1881 PHINode *ResumeIndex = 0;
1882 LoopVectorizationLegality::InductionList::iterator I, E;
1883 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1884 // Set builder to point to last bypass block.
1885 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1886 for (I = List->begin(), E = List->end(); I != E; ++I) {
1887 PHINode *OrigPhi = I->first;
1888 LoopVectorizationLegality::InductionInfo II = I->second;
1890 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1891 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1892 MiddleBlock->getTerminator());
1893 // We might have extended the type of the induction variable but we need a
1894 // truncated version for the scalar loop.
1895 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1896 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1897 MiddleBlock->getTerminator()) : 0;
1899 Value *EndValue = 0;
1901 case LoopVectorizationLegality::IK_NoInduction:
1902 llvm_unreachable("Unknown induction");
1903 case LoopVectorizationLegality::IK_IntInduction: {
1904 // Handle the integer induction counter.
1905 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1907 // We have the canonical induction variable.
1908 if (OrigPhi == OldInduction) {
1909 // Create a truncated version of the resume value for the scalar loop,
1910 // we might have promoted the type to a larger width.
1912 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1913 // The new PHI merges the original incoming value, in case of a bypass,
1914 // or the value at the end of the vectorized loop.
1915 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1916 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1917 TruncResumeVal->addIncoming(EndValue, VecBody);
1919 // We know what the end value is.
1920 EndValue = IdxEndRoundDown;
1921 // We also know which PHI node holds it.
1922 ResumeIndex = ResumeVal;
1926 // Not the canonical induction variable - add the vector loop count to the
1928 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1929 II.StartValue->getType(),
1931 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1934 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1935 // Convert the CountRoundDown variable to the PHI size.
1936 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1937 II.StartValue->getType(),
1939 // Handle reverse integer induction counter.
1940 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1943 case LoopVectorizationLegality::IK_PtrInduction: {
1944 // For pointer induction variables, calculate the offset using
1946 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1950 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1951 // The value at the end of the loop for the reverse pointer is calculated
1952 // by creating a GEP with a negative index starting from the start value.
1953 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1954 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1956 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1962 // The new PHI merges the original incoming value, in case of a bypass,
1963 // or the value at the end of the vectorized loop.
1964 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1965 if (OrigPhi == OldInduction)
1966 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1968 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1970 ResumeVal->addIncoming(EndValue, VecBody);
1972 // Fix the scalar body counter (PHI node).
1973 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1974 // The old inductions phi node in the scalar body needs the truncated value.
1975 if (OrigPhi == OldInduction)
1976 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1978 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1981 // If we are generating a new induction variable then we also need to
1982 // generate the code that calculates the exit value. This value is not
1983 // simply the end of the counter because we may skip the vectorized body
1984 // in case of a runtime check.
1986 assert(!ResumeIndex && "Unexpected resume value found");
1987 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1988 MiddleBlock->getTerminator());
1989 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1990 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1991 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1994 // Make sure that we found the index where scalar loop needs to continue.
1995 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1996 "Invalid resume Index");
1998 // Add a check in the middle block to see if we have completed
1999 // all of the iterations in the first vector loop.
2000 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2001 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2002 ResumeIndex, "cmp.n",
2003 MiddleBlock->getTerminator());
2005 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2006 // Remove the old terminator.
2007 MiddleBlock->getTerminator()->eraseFromParent();
2009 // Create i+1 and fill the PHINode.
2010 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2011 Induction->addIncoming(StartIdx, VectorPH);
2012 Induction->addIncoming(NextIdx, VecBody);
2013 // Create the compare.
2014 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2015 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2017 // Now we have two terminators. Remove the old one from the block.
2018 VecBody->getTerminator()->eraseFromParent();
2020 // Get ready to start creating new instructions into the vectorized body.
2021 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2024 LoopVectorPreHeader = VectorPH;
2025 LoopScalarPreHeader = ScalarPH;
2026 LoopMiddleBlock = MiddleBlock;
2027 LoopExitBlock = ExitBlock;
2028 LoopVectorBody = VecBody;
2029 LoopScalarBody = OldBasicBlock;
2031 LoopVectorizeHints Hints(Lp, true);
2032 Hints.setAlreadyVectorized(Lp);
2035 /// This function returns the identity element (or neutral element) for
2036 /// the operation K.
2038 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2043 // Adding, Xoring, Oring zero to a number does not change it.
2044 return ConstantInt::get(Tp, 0);
2045 case RK_IntegerMult:
2046 // Multiplying a number by 1 does not change it.
2047 return ConstantInt::get(Tp, 1);
2049 // AND-ing a number with an all-1 value does not change it.
2050 return ConstantInt::get(Tp, -1, true);
2052 // Multiplying a number by 1 does not change it.
2053 return ConstantFP::get(Tp, 1.0L);
2055 // Adding zero to a number does not change it.
2056 return ConstantFP::get(Tp, 0.0L);
2058 llvm_unreachable("Unknown reduction kind");
2062 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2063 Intrinsic::ID ValidIntrinsicID) {
2064 if (I.getNumArgOperands() != 1 ||
2065 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2066 I.getType() != I.getArgOperand(0)->getType() ||
2067 !I.onlyReadsMemory())
2068 return Intrinsic::not_intrinsic;
2070 return ValidIntrinsicID;
2073 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2074 Intrinsic::ID ValidIntrinsicID) {
2075 if (I.getNumArgOperands() != 2 ||
2076 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2077 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2078 I.getType() != I.getArgOperand(0)->getType() ||
2079 I.getType() != I.getArgOperand(1)->getType() ||
2080 !I.onlyReadsMemory())
2081 return Intrinsic::not_intrinsic;
2083 return ValidIntrinsicID;
2087 static Intrinsic::ID
2088 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2089 // If we have an intrinsic call, check if it is trivially vectorizable.
2090 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2091 switch (II->getIntrinsicID()) {
2092 case Intrinsic::sqrt:
2093 case Intrinsic::sin:
2094 case Intrinsic::cos:
2095 case Intrinsic::exp:
2096 case Intrinsic::exp2:
2097 case Intrinsic::log:
2098 case Intrinsic::log10:
2099 case Intrinsic::log2:
2100 case Intrinsic::fabs:
2101 case Intrinsic::copysign:
2102 case Intrinsic::floor:
2103 case Intrinsic::ceil:
2104 case Intrinsic::trunc:
2105 case Intrinsic::rint:
2106 case Intrinsic::nearbyint:
2107 case Intrinsic::round:
2108 case Intrinsic::pow:
2109 case Intrinsic::fma:
2110 case Intrinsic::fmuladd:
2111 case Intrinsic::lifetime_start:
2112 case Intrinsic::lifetime_end:
2113 return II->getIntrinsicID();
2115 return Intrinsic::not_intrinsic;
2120 return Intrinsic::not_intrinsic;
2123 Function *F = CI->getCalledFunction();
2124 // We're going to make assumptions on the semantics of the functions, check
2125 // that the target knows that it's available in this environment and it does
2126 // not have local linkage.
2127 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2128 return Intrinsic::not_intrinsic;
2130 // Otherwise check if we have a call to a function that can be turned into a
2131 // vector intrinsic.
2138 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2142 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2146 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2148 case LibFunc::exp2f:
2149 case LibFunc::exp2l:
2150 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2154 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2155 case LibFunc::log10:
2156 case LibFunc::log10f:
2157 case LibFunc::log10l:
2158 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2160 case LibFunc::log2f:
2161 case LibFunc::log2l:
2162 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2164 case LibFunc::fabsf:
2165 case LibFunc::fabsl:
2166 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2167 case LibFunc::copysign:
2168 case LibFunc::copysignf:
2169 case LibFunc::copysignl:
2170 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2171 case LibFunc::floor:
2172 case LibFunc::floorf:
2173 case LibFunc::floorl:
2174 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2176 case LibFunc::ceilf:
2177 case LibFunc::ceill:
2178 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2179 case LibFunc::trunc:
2180 case LibFunc::truncf:
2181 case LibFunc::truncl:
2182 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2184 case LibFunc::rintf:
2185 case LibFunc::rintl:
2186 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2187 case LibFunc::nearbyint:
2188 case LibFunc::nearbyintf:
2189 case LibFunc::nearbyintl:
2190 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2191 case LibFunc::round:
2192 case LibFunc::roundf:
2193 case LibFunc::roundl:
2194 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2198 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2201 return Intrinsic::not_intrinsic;
2204 /// This function translates the reduction kind to an LLVM binary operator.
2206 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2208 case LoopVectorizationLegality::RK_IntegerAdd:
2209 return Instruction::Add;
2210 case LoopVectorizationLegality::RK_IntegerMult:
2211 return Instruction::Mul;
2212 case LoopVectorizationLegality::RK_IntegerOr:
2213 return Instruction::Or;
2214 case LoopVectorizationLegality::RK_IntegerAnd:
2215 return Instruction::And;
2216 case LoopVectorizationLegality::RK_IntegerXor:
2217 return Instruction::Xor;
2218 case LoopVectorizationLegality::RK_FloatMult:
2219 return Instruction::FMul;
2220 case LoopVectorizationLegality::RK_FloatAdd:
2221 return Instruction::FAdd;
2222 case LoopVectorizationLegality::RK_IntegerMinMax:
2223 return Instruction::ICmp;
2224 case LoopVectorizationLegality::RK_FloatMinMax:
2225 return Instruction::FCmp;
2227 llvm_unreachable("Unknown reduction operation");
2231 Value *createMinMaxOp(IRBuilder<> &Builder,
2232 LoopVectorizationLegality::MinMaxReductionKind RK,
2235 CmpInst::Predicate P = CmpInst::ICMP_NE;
2238 llvm_unreachable("Unknown min/max reduction kind");
2239 case LoopVectorizationLegality::MRK_UIntMin:
2240 P = CmpInst::ICMP_ULT;
2242 case LoopVectorizationLegality::MRK_UIntMax:
2243 P = CmpInst::ICMP_UGT;
2245 case LoopVectorizationLegality::MRK_SIntMin:
2246 P = CmpInst::ICMP_SLT;
2248 case LoopVectorizationLegality::MRK_SIntMax:
2249 P = CmpInst::ICMP_SGT;
2251 case LoopVectorizationLegality::MRK_FloatMin:
2252 P = CmpInst::FCMP_OLT;
2254 case LoopVectorizationLegality::MRK_FloatMax:
2255 P = CmpInst::FCMP_OGT;
2260 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2261 RK == LoopVectorizationLegality::MRK_FloatMax)
2262 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2264 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2266 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2271 struct CSEDenseMapInfo {
2272 static bool canHandle(Instruction *I) {
2273 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2274 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2276 static inline Instruction *getEmptyKey() {
2277 return DenseMapInfo<Instruction *>::getEmptyKey();
2279 static inline Instruction *getTombstoneKey() {
2280 return DenseMapInfo<Instruction *>::getTombstoneKey();
2282 static unsigned getHashValue(Instruction *I) {
2283 assert(canHandle(I) && "Unknown instruction!");
2284 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2285 I->value_op_end()));
2287 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2288 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2289 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2291 return LHS->isIdenticalTo(RHS);
2296 ///\brief Perform cse of induction variable instructions.
2297 static void cse(BasicBlock *BB) {
2298 // Perform simple cse.
2299 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2300 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2301 Instruction *In = I++;
2303 if (!CSEDenseMapInfo::canHandle(In))
2306 // Check if we can replace this instruction with any of the
2307 // visited instructions.
2308 if (Instruction *V = CSEMap.lookup(In)) {
2309 In->replaceAllUsesWith(V);
2310 In->eraseFromParent();
2318 void InnerLoopVectorizer::vectorizeLoop() {
2319 //===------------------------------------------------===//
2321 // Notice: any optimization or new instruction that go
2322 // into the code below should be also be implemented in
2325 //===------------------------------------------------===//
2326 Constant *Zero = Builder.getInt32(0);
2328 // In order to support reduction variables we need to be able to vectorize
2329 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2330 // stages. First, we create a new vector PHI node with no incoming edges.
2331 // We use this value when we vectorize all of the instructions that use the
2332 // PHI. Next, after all of the instructions in the block are complete we
2333 // add the new incoming edges to the PHI. At this point all of the
2334 // instructions in the basic block are vectorized, so we can use them to
2335 // construct the PHI.
2336 PhiVector RdxPHIsToFix;
2338 // Scan the loop in a topological order to ensure that defs are vectorized
2340 LoopBlocksDFS DFS(OrigLoop);
2343 // Vectorize all of the blocks in the original loop.
2344 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2345 be = DFS.endRPO(); bb != be; ++bb)
2346 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2348 // At this point every instruction in the original loop is widened to
2349 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2350 // that we vectorized. The PHI nodes are currently empty because we did
2351 // not want to introduce cycles. Notice that the remaining PHI nodes
2352 // that we need to fix are reduction variables.
2354 // Create the 'reduced' values for each of the induction vars.
2355 // The reduced values are the vector values that we scalarize and combine
2356 // after the loop is finished.
2357 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2359 PHINode *RdxPhi = *it;
2360 assert(RdxPhi && "Unable to recover vectorized PHI");
2362 // Find the reduction variable descriptor.
2363 assert(Legal->getReductionVars()->count(RdxPhi) &&
2364 "Unable to find the reduction variable");
2365 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2366 (*Legal->getReductionVars())[RdxPhi];
2368 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2370 // We need to generate a reduction vector from the incoming scalar.
2371 // To do so, we need to generate the 'identity' vector and overide
2372 // one of the elements with the incoming scalar reduction. We need
2373 // to do it in the vector-loop preheader.
2374 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2376 // This is the vector-clone of the value that leaves the loop.
2377 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2378 Type *VecTy = VectorExit[0]->getType();
2380 // Find the reduction identity variable. Zero for addition, or, xor,
2381 // one for multiplication, -1 for And.
2384 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2385 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2386 // MinMax reduction have the start value as their identify.
2388 VectorStart = Identity = RdxDesc.StartValue;
2390 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2395 // Handle other reduction kinds:
2397 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2398 VecTy->getScalarType());
2401 // This vector is the Identity vector where the first element is the
2402 // incoming scalar reduction.
2403 VectorStart = RdxDesc.StartValue;
2405 Identity = ConstantVector::getSplat(VF, Iden);
2407 // This vector is the Identity vector where the first element is the
2408 // incoming scalar reduction.
2409 VectorStart = Builder.CreateInsertElement(Identity,
2410 RdxDesc.StartValue, Zero);
2414 // Fix the vector-loop phi.
2415 // We created the induction variable so we know that the
2416 // preheader is the first entry.
2417 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2419 // Reductions do not have to start at zero. They can start with
2420 // any loop invariant values.
2421 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2422 BasicBlock *Latch = OrigLoop->getLoopLatch();
2423 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2424 VectorParts &Val = getVectorValue(LoopVal);
2425 for (unsigned part = 0; part < UF; ++part) {
2426 // Make sure to add the reduction stat value only to the
2427 // first unroll part.
2428 Value *StartVal = (part == 0) ? VectorStart : Identity;
2429 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2430 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2433 // Before each round, move the insertion point right between
2434 // the PHIs and the values we are going to write.
2435 // This allows us to write both PHINodes and the extractelement
2437 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2439 VectorParts RdxParts;
2440 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2441 for (unsigned part = 0; part < UF; ++part) {
2442 // This PHINode contains the vectorized reduction variable, or
2443 // the initial value vector, if we bypass the vector loop.
2444 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2445 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2446 Value *StartVal = (part == 0) ? VectorStart : Identity;
2447 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2448 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2449 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2450 RdxParts.push_back(NewPhi);
2453 // Reduce all of the unrolled parts into a single vector.
2454 Value *ReducedPartRdx = RdxParts[0];
2455 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2456 setDebugLocFromInst(Builder, ReducedPartRdx);
2457 for (unsigned part = 1; part < UF; ++part) {
2458 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2459 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2460 RdxParts[part], ReducedPartRdx,
2463 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2464 ReducedPartRdx, RdxParts[part]);
2468 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2469 // and vector ops, reducing the set of values being computed by half each
2471 assert(isPowerOf2_32(VF) &&
2472 "Reduction emission only supported for pow2 vectors!");
2473 Value *TmpVec = ReducedPartRdx;
2474 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2475 for (unsigned i = VF; i != 1; i >>= 1) {
2476 // Move the upper half of the vector to the lower half.
2477 for (unsigned j = 0; j != i/2; ++j)
2478 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2480 // Fill the rest of the mask with undef.
2481 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2482 UndefValue::get(Builder.getInt32Ty()));
2485 Builder.CreateShuffleVector(TmpVec,
2486 UndefValue::get(TmpVec->getType()),
2487 ConstantVector::get(ShuffleMask),
2490 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2491 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2494 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2497 // The result is in the first element of the vector.
2498 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2499 Builder.getInt32(0));
2502 // Now, we need to fix the users of the reduction variable
2503 // inside and outside of the scalar remainder loop.
2504 // We know that the loop is in LCSSA form. We need to update the
2505 // PHI nodes in the exit blocks.
2506 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2507 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2508 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2509 if (!LCSSAPhi) break;
2511 // All PHINodes need to have a single entry edge, or two if
2512 // we already fixed them.
2513 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2515 // We found our reduction value exit-PHI. Update it with the
2516 // incoming bypass edge.
2517 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2518 // Add an edge coming from the bypass.
2519 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2522 }// end of the LCSSA phi scan.
2524 // Fix the scalar loop reduction variable with the incoming reduction sum
2525 // from the vector body and from the backedge value.
2526 int IncomingEdgeBlockIdx =
2527 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2528 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2529 // Pick the other block.
2530 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2531 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2532 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2533 }// end of for each redux variable.
2537 // Remove redundant induction instructions.
2538 cse(LoopVectorBody);
2541 void InnerLoopVectorizer::fixLCSSAPHIs() {
2542 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2543 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2544 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2545 if (!LCSSAPhi) break;
2546 if (LCSSAPhi->getNumIncomingValues() == 1)
2547 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2552 InnerLoopVectorizer::VectorParts
2553 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2554 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2557 // Look for cached value.
2558 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2559 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2560 if (ECEntryIt != MaskCache.end())
2561 return ECEntryIt->second;
2563 VectorParts SrcMask = createBlockInMask(Src);
2565 // The terminator has to be a branch inst!
2566 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2567 assert(BI && "Unexpected terminator found");
2569 if (BI->isConditional()) {
2570 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2572 if (BI->getSuccessor(0) != Dst)
2573 for (unsigned part = 0; part < UF; ++part)
2574 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2576 for (unsigned part = 0; part < UF; ++part)
2577 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2579 MaskCache[Edge] = EdgeMask;
2583 MaskCache[Edge] = SrcMask;
2587 InnerLoopVectorizer::VectorParts
2588 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2589 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2591 // Loop incoming mask is all-one.
2592 if (OrigLoop->getHeader() == BB) {
2593 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2594 return getVectorValue(C);
2597 // This is the block mask. We OR all incoming edges, and with zero.
2598 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2599 VectorParts BlockMask = getVectorValue(Zero);
2602 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2603 VectorParts EM = createEdgeMask(*it, BB);
2604 for (unsigned part = 0; part < UF; ++part)
2605 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2611 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2612 InnerLoopVectorizer::VectorParts &Entry,
2613 unsigned UF, unsigned VF, PhiVector *PV) {
2614 PHINode* P = cast<PHINode>(PN);
2615 // Handle reduction variables:
2616 if (Legal->getReductionVars()->count(P)) {
2617 for (unsigned part = 0; part < UF; ++part) {
2618 // This is phase one of vectorizing PHIs.
2619 Type *VecTy = (VF == 1) ? PN->getType() :
2620 VectorType::get(PN->getType(), VF);
2621 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2622 LoopVectorBody-> getFirstInsertionPt());
2628 setDebugLocFromInst(Builder, P);
2629 // Check for PHI nodes that are lowered to vector selects.
2630 if (P->getParent() != OrigLoop->getHeader()) {
2631 // We know that all PHIs in non-header blocks are converted into
2632 // selects, so we don't have to worry about the insertion order and we
2633 // can just use the builder.
2634 // At this point we generate the predication tree. There may be
2635 // duplications since this is a simple recursive scan, but future
2636 // optimizations will clean it up.
2638 unsigned NumIncoming = P->getNumIncomingValues();
2640 // Generate a sequence of selects of the form:
2641 // SELECT(Mask3, In3,
2642 // SELECT(Mask2, In2,
2644 for (unsigned In = 0; In < NumIncoming; In++) {
2645 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2647 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2649 for (unsigned part = 0; part < UF; ++part) {
2650 // We might have single edge PHIs (blocks) - use an identity
2651 // 'select' for the first PHI operand.
2653 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2656 // Select between the current value and the previous incoming edge
2657 // based on the incoming mask.
2658 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2659 Entry[part], "predphi");
2665 // This PHINode must be an induction variable.
2666 // Make sure that we know about it.
2667 assert(Legal->getInductionVars()->count(P) &&
2668 "Not an induction variable");
2670 LoopVectorizationLegality::InductionInfo II =
2671 Legal->getInductionVars()->lookup(P);
2674 case LoopVectorizationLegality::IK_NoInduction:
2675 llvm_unreachable("Unknown induction");
2676 case LoopVectorizationLegality::IK_IntInduction: {
2677 assert(P->getType() == II.StartValue->getType() && "Types must match");
2678 Type *PhiTy = P->getType();
2680 if (P == OldInduction) {
2681 // Handle the canonical induction variable. We might have had to
2683 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2685 // Handle other induction variables that are now based on the
2687 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2689 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2690 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2693 Broadcasted = getBroadcastInstrs(Broadcasted);
2694 // After broadcasting the induction variable we need to make the vector
2695 // consecutive by adding 0, 1, 2, etc.
2696 for (unsigned part = 0; part < UF; ++part)
2697 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2700 case LoopVectorizationLegality::IK_ReverseIntInduction:
2701 case LoopVectorizationLegality::IK_PtrInduction:
2702 case LoopVectorizationLegality::IK_ReversePtrInduction:
2703 // Handle reverse integer and pointer inductions.
2704 Value *StartIdx = ExtendedIdx;
2705 // This is the normalized GEP that starts counting at zero.
2706 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2709 // Handle the reverse integer induction variable case.
2710 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2711 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2712 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2714 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2717 // This is a new value so do not hoist it out.
2718 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2719 // After broadcasting the induction variable we need to make the
2720 // vector consecutive by adding ... -3, -2, -1, 0.
2721 for (unsigned part = 0; part < UF; ++part)
2722 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2727 // Handle the pointer induction variable case.
2728 assert(P->getType()->isPointerTy() && "Unexpected type.");
2730 // Is this a reverse induction ptr or a consecutive induction ptr.
2731 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2734 // This is the vector of results. Notice that we don't generate
2735 // vector geps because scalar geps result in better code.
2736 for (unsigned part = 0; part < UF; ++part) {
2738 int EltIndex = (part) * (Reverse ? -1 : 1);
2739 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2742 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2744 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2746 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2748 Entry[part] = SclrGep;
2752 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2753 for (unsigned int i = 0; i < VF; ++i) {
2754 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2755 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2758 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2760 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2762 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2764 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2765 Builder.getInt32(i),
2768 Entry[part] = VecVal;
2774 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2775 // For each instruction in the old loop.
2776 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2777 VectorParts &Entry = WidenMap.get(it);
2778 switch (it->getOpcode()) {
2779 case Instruction::Br:
2780 // Nothing to do for PHIs and BR, since we already took care of the
2781 // loop control flow instructions.
2783 case Instruction::PHI:{
2784 // Vectorize PHINodes.
2785 widenPHIInstruction(it, Entry, UF, VF, PV);
2789 case Instruction::Add:
2790 case Instruction::FAdd:
2791 case Instruction::Sub:
2792 case Instruction::FSub:
2793 case Instruction::Mul:
2794 case Instruction::FMul:
2795 case Instruction::UDiv:
2796 case Instruction::SDiv:
2797 case Instruction::FDiv:
2798 case Instruction::URem:
2799 case Instruction::SRem:
2800 case Instruction::FRem:
2801 case Instruction::Shl:
2802 case Instruction::LShr:
2803 case Instruction::AShr:
2804 case Instruction::And:
2805 case Instruction::Or:
2806 case Instruction::Xor: {
2807 // Just widen binops.
2808 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2809 setDebugLocFromInst(Builder, BinOp);
2810 VectorParts &A = getVectorValue(it->getOperand(0));
2811 VectorParts &B = getVectorValue(it->getOperand(1));
2813 // Use this vector value for all users of the original instruction.
2814 for (unsigned Part = 0; Part < UF; ++Part) {
2815 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2817 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2818 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2819 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2820 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2821 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2823 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2824 VecOp->setIsExact(BinOp->isExact());
2830 case Instruction::Select: {
2832 // If the selector is loop invariant we can create a select
2833 // instruction with a scalar condition. Otherwise, use vector-select.
2834 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2836 setDebugLocFromInst(Builder, it);
2838 // The condition can be loop invariant but still defined inside the
2839 // loop. This means that we can't just use the original 'cond' value.
2840 // We have to take the 'vectorized' value and pick the first lane.
2841 // Instcombine will make this a no-op.
2842 VectorParts &Cond = getVectorValue(it->getOperand(0));
2843 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2844 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2846 Value *ScalarCond = (VF == 1) ? Cond[0] :
2847 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2849 for (unsigned Part = 0; Part < UF; ++Part) {
2850 Entry[Part] = Builder.CreateSelect(
2851 InvariantCond ? ScalarCond : Cond[Part],
2858 case Instruction::ICmp:
2859 case Instruction::FCmp: {
2860 // Widen compares. Generate vector compares.
2861 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2862 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2863 setDebugLocFromInst(Builder, it);
2864 VectorParts &A = getVectorValue(it->getOperand(0));
2865 VectorParts &B = getVectorValue(it->getOperand(1));
2866 for (unsigned Part = 0; Part < UF; ++Part) {
2869 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2871 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2877 case Instruction::Store:
2878 case Instruction::Load:
2879 vectorizeMemoryInstruction(it);
2881 case Instruction::ZExt:
2882 case Instruction::SExt:
2883 case Instruction::FPToUI:
2884 case Instruction::FPToSI:
2885 case Instruction::FPExt:
2886 case Instruction::PtrToInt:
2887 case Instruction::IntToPtr:
2888 case Instruction::SIToFP:
2889 case Instruction::UIToFP:
2890 case Instruction::Trunc:
2891 case Instruction::FPTrunc:
2892 case Instruction::BitCast: {
2893 CastInst *CI = dyn_cast<CastInst>(it);
2894 setDebugLocFromInst(Builder, it);
2895 /// Optimize the special case where the source is the induction
2896 /// variable. Notice that we can only optimize the 'trunc' case
2897 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2898 /// c. other casts depend on pointer size.
2899 if (CI->getOperand(0) == OldInduction &&
2900 it->getOpcode() == Instruction::Trunc) {
2901 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2903 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2904 for (unsigned Part = 0; Part < UF; ++Part)
2905 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2908 /// Vectorize casts.
2909 Type *DestTy = (VF == 1) ? CI->getType() :
2910 VectorType::get(CI->getType(), VF);
2912 VectorParts &A = getVectorValue(it->getOperand(0));
2913 for (unsigned Part = 0; Part < UF; ++Part)
2914 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2918 case Instruction::Call: {
2919 // Ignore dbg intrinsics.
2920 if (isa<DbgInfoIntrinsic>(it))
2922 setDebugLocFromInst(Builder, it);
2924 Module *M = BB->getParent()->getParent();
2925 CallInst *CI = cast<CallInst>(it);
2926 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2927 assert(ID && "Not an intrinsic call!");
2929 case Intrinsic::lifetime_end:
2930 case Intrinsic::lifetime_start:
2931 scalarizeInstruction(it);
2934 for (unsigned Part = 0; Part < UF; ++Part) {
2935 SmallVector<Value *, 4> Args;
2936 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2937 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2938 Args.push_back(Arg[Part]);
2940 Type *Tys[] = {CI->getType()};
2942 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2944 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2945 Entry[Part] = Builder.CreateCall(F, Args);
2953 // All other instructions are unsupported. Scalarize them.
2954 scalarizeInstruction(it);
2957 }// end of for_each instr.
2960 void InnerLoopVectorizer::updateAnalysis() {
2961 // Forget the original basic block.
2962 SE->forgetLoop(OrigLoop);
2964 // Update the dominator tree information.
2965 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2966 "Entry does not dominate exit.");
2968 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2969 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2970 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2971 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2972 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2973 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2974 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2975 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2977 DEBUG(DT->verifyAnalysis());
2980 /// \brief Check whether it is safe to if-convert this phi node.
2982 /// Phi nodes with constant expressions that can trap are not safe to if
2984 static bool canIfConvertPHINodes(BasicBlock *BB) {
2985 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2986 PHINode *Phi = dyn_cast<PHINode>(I);
2989 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
2990 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
2997 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2998 if (!EnableIfConversion)
3001 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3003 // A list of pointers that we can safely read and write to.
3004 SmallPtrSet<Value *, 8> SafePointes;
3006 // Collect safe addresses.
3007 for (Loop::block_iterator BI = TheLoop->block_begin(),
3008 BE = TheLoop->block_end(); BI != BE; ++BI) {
3009 BasicBlock *BB = *BI;
3011 if (blockNeedsPredication(BB))
3014 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3015 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3016 SafePointes.insert(LI->getPointerOperand());
3017 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3018 SafePointes.insert(SI->getPointerOperand());
3022 // Collect the blocks that need predication.
3023 BasicBlock *Header = TheLoop->getHeader();
3024 for (Loop::block_iterator BI = TheLoop->block_begin(),
3025 BE = TheLoop->block_end(); BI != BE; ++BI) {
3026 BasicBlock *BB = *BI;
3028 // We don't support switch statements inside loops.
3029 if (!isa<BranchInst>(BB->getTerminator()))
3032 // We must be able to predicate all blocks that need to be predicated.
3033 if (blockNeedsPredication(BB)) {
3034 if (!blockCanBePredicated(BB, SafePointes))
3036 } else if (BB != Header && !canIfConvertPHINodes(BB))
3041 // We can if-convert this loop.
3045 bool LoopVectorizationLegality::canVectorize() {
3046 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3047 // be canonicalized.
3048 if (!TheLoop->getLoopPreheader())
3051 // We can only vectorize innermost loops.
3052 if (TheLoop->getSubLoopsVector().size())
3055 // We must have a single backedge.
3056 if (TheLoop->getNumBackEdges() != 1)
3059 // We must have a single exiting block.
3060 if (!TheLoop->getExitingBlock())
3063 // We need to have a loop header.
3064 DEBUG(dbgs() << "LV: Found a loop: " <<
3065 TheLoop->getHeader()->getName() << '\n');
3067 // Check if we can if-convert non-single-bb loops.
3068 unsigned NumBlocks = TheLoop->getNumBlocks();
3069 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3070 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3074 // ScalarEvolution needs to be able to find the exit count.
3075 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3076 if (ExitCount == SE->getCouldNotCompute()) {
3077 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3081 // Do not loop-vectorize loops with a tiny trip count.
3082 BasicBlock *Latch = TheLoop->getLoopLatch();
3083 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
3084 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
3085 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
3086 "This loop is not worth vectorizing.\n");
3090 // Check if we can vectorize the instructions and CFG in this loop.
3091 if (!canVectorizeInstrs()) {
3092 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3096 // Go over each instruction and look at memory deps.
3097 if (!canVectorizeMemory()) {
3098 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3102 // Collect all of the variables that remain uniform after vectorization.
3103 collectLoopUniforms();
3105 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3106 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3109 // Okay! We can vectorize. At this point we don't have any other mem analysis
3110 // which may limit our maximum vectorization factor, so just return true with
3115 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
3116 if (Ty->isPointerTy())
3117 return DL.getIntPtrType(Ty);
3119 // It is possible that char's or short's overflow when we ask for the loop's
3120 // trip count, work around this by changing the type size.
3121 if (Ty->getScalarSizeInBits() < 32)
3122 return Type::getInt32Ty(Ty->getContext());
3127 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
3128 Ty0 = convertPointerToIntegerType(DL, Ty0);
3129 Ty1 = convertPointerToIntegerType(DL, Ty1);
3130 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3135 /// \brief Check that the instruction has outside loop users and is not an
3136 /// identified reduction variable.
3137 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3138 SmallPtrSet<Value *, 4> &Reductions) {
3139 // Reduction instructions are allowed to have exit users. All other
3140 // instructions must not have external users.
3141 if (!Reductions.count(Inst))
3142 //Check that all of the users of the loop are inside the BB.
3143 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
3145 Instruction *U = cast<Instruction>(*I);
3146 // This user may be a reduction exit value.
3147 if (!TheLoop->contains(U)) {
3148 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
3155 bool LoopVectorizationLegality::canVectorizeInstrs() {
3156 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3157 BasicBlock *Header = TheLoop->getHeader();
3159 // Look for the attribute signaling the absence of NaNs.
3160 Function &F = *Header->getParent();
3161 if (F.hasFnAttribute("no-nans-fp-math"))
3162 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3163 AttributeSet::FunctionIndex,
3164 "no-nans-fp-math").getValueAsString() == "true";
3166 // For each block in the loop.
3167 for (Loop::block_iterator bb = TheLoop->block_begin(),
3168 be = TheLoop->block_end(); bb != be; ++bb) {
3170 // Scan the instructions in the block and look for hazards.
3171 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3174 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3175 Type *PhiTy = Phi->getType();
3176 // Check that this PHI type is allowed.
3177 if (!PhiTy->isIntegerTy() &&
3178 !PhiTy->isFloatingPointTy() &&
3179 !PhiTy->isPointerTy()) {
3180 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3184 // If this PHINode is not in the header block, then we know that we
3185 // can convert it to select during if-conversion. No need to check if
3186 // the PHIs in this block are induction or reduction variables.
3187 if (*bb != Header) {
3188 // Check that this instruction has no outside users or is an
3189 // identified reduction value with an outside user.
3190 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3195 // We only allow if-converted PHIs with more than two incoming values.
3196 if (Phi->getNumIncomingValues() != 2) {
3197 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3201 // This is the value coming from the preheader.
3202 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3203 // Check if this is an induction variable.
3204 InductionKind IK = isInductionVariable(Phi);
3206 if (IK_NoInduction != IK) {
3207 // Get the widest type.
3209 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3211 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3213 // Int inductions are special because we only allow one IV.
3214 if (IK == IK_IntInduction) {
3215 // Use the phi node with the widest type as induction. Use the last
3216 // one if there are multiple (no good reason for doing this other
3217 // than it is expedient).
3218 if (!Induction || PhiTy == WidestIndTy)
3222 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3223 Inductions[Phi] = InductionInfo(StartValue, IK);
3225 // Until we explicitly handle the case of an induction variable with
3226 // an outside loop user we have to give up vectorizing this loop.
3227 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3233 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3234 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3237 if (AddReductionVar(Phi, RK_IntegerMult)) {
3238 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3241 if (AddReductionVar(Phi, RK_IntegerOr)) {
3242 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3245 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3246 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3249 if (AddReductionVar(Phi, RK_IntegerXor)) {
3250 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3253 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3254 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3257 if (AddReductionVar(Phi, RK_FloatMult)) {
3258 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3261 if (AddReductionVar(Phi, RK_FloatAdd)) {
3262 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3265 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3266 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3271 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3273 }// end of PHI handling
3275 // We still don't handle functions. However, we can ignore dbg intrinsic
3276 // calls and we do handle certain intrinsic and libm functions.
3277 CallInst *CI = dyn_cast<CallInst>(it);
3278 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3279 DEBUG(dbgs() << "LV: Found a call site.\n");
3283 // Check that the instruction return type is vectorizable.
3284 // Also, we can't vectorize extractelement instructions.
3285 if ((!VectorType::isValidElementType(it->getType()) &&
3286 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3287 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3291 // Check that the stored type is vectorizable.
3292 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3293 Type *T = ST->getValueOperand()->getType();
3294 if (!VectorType::isValidElementType(T))
3296 if (EnableMemAccessVersioning)
3297 collectStridedAcccess(ST);
3300 if (EnableMemAccessVersioning)
3301 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3302 collectStridedAcccess(LI);
3304 // Reduction instructions are allowed to have exit users.
3305 // All other instructions must not have external users.
3306 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3314 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3315 if (Inductions.empty())
3322 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3323 /// return the induction operand of the gep pointer.
3324 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3325 DataLayout *DL, Loop *Lp) {
3326 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3330 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3332 // Check that all of the gep indices are uniform except for our induction
3334 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3335 if (i != InductionOperand &&
3336 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3338 return GEP->getOperand(InductionOperand);
3341 ///\brief Look for a cast use of the passed value.
3342 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3343 Value *UniqueCast = 0;
3344 for (Value::use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end(); UI != UE;
3346 CastInst *CI = dyn_cast<CastInst>(*UI);
3347 if (CI && CI->getType() == Ty) {
3357 ///\brief Get the stride of a pointer access in a loop.
3358 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3359 /// pointer to the Value, or null otherwise.
3360 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3361 DataLayout *DL, Loop *Lp) {
3362 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3363 if (!PtrTy || PtrTy->isAggregateType())
3366 // Try to remove a gep instruction to make the pointer (actually index at this
3367 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3368 // pointer, otherwise, we are analyzing the index.
3369 Value *OrigPtr = Ptr;
3371 // The size of the pointer access.
3372 int64_t PtrAccessSize = 1;
3374 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3375 const SCEV *V = SE->getSCEV(Ptr);
3379 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3380 V = C->getOperand();
3382 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3386 V = S->getStepRecurrence(*SE);
3390 // Strip off the size of access multiplication if we are still analyzing the
3392 if (OrigPtr == Ptr) {
3393 DL->getTypeAllocSize(PtrTy->getElementType());
3394 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3395 if (M->getOperand(0)->getSCEVType() != scConstant)
3398 const APInt &APStepVal =
3399 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3401 // Huge step value - give up.
3402 if (APStepVal.getBitWidth() > 64)
3405 int64_t StepVal = APStepVal.getSExtValue();
3406 if (PtrAccessSize != StepVal)
3408 V = M->getOperand(1);
3413 Type *StripedOffRecurrenceCast = 0;
3414 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3415 StripedOffRecurrenceCast = C->getType();
3416 V = C->getOperand();
3419 // Look for the loop invariant symbolic value.
3420 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3424 Value *Stride = U->getValue();
3425 if (!Lp->isLoopInvariant(Stride))
3428 // If we have stripped off the recurrence cast we have to make sure that we
3429 // return the value that is used in this loop so that we can replace it later.
3430 if (StripedOffRecurrenceCast)
3431 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3436 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3438 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3439 Ptr = LI->getPointerOperand();
3440 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3441 Ptr = SI->getPointerOperand();
3445 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3449 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3450 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3451 Strides[Ptr] = Stride;
3452 StrideSet.insert(Stride);
3455 void LoopVectorizationLegality::collectLoopUniforms() {
3456 // We now know that the loop is vectorizable!
3457 // Collect variables that will remain uniform after vectorization.
3458 std::vector<Value*> Worklist;
3459 BasicBlock *Latch = TheLoop->getLoopLatch();
3461 // Start with the conditional branch and walk up the block.
3462 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3464 while (Worklist.size()) {
3465 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3466 Worklist.pop_back();
3468 // Look at instructions inside this loop.
3469 // Stop when reaching PHI nodes.
3470 // TODO: we need to follow values all over the loop, not only in this block.
3471 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3474 // This is a known uniform.
3477 // Insert all operands.
3478 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3483 /// \brief Analyses memory accesses in a loop.
3485 /// Checks whether run time pointer checks are needed and builds sets for data
3486 /// dependence checking.
3487 class AccessAnalysis {
3489 /// \brief Read or write access location.
3490 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3491 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3493 /// \brief Set of potential dependent memory accesses.
3494 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3496 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3497 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3498 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3500 /// \brief Register a load and whether it is only read from.
3501 void addLoad(Value *Ptr, bool IsReadOnly) {
3502 Accesses.insert(MemAccessInfo(Ptr, false));
3504 ReadOnlyPtr.insert(Ptr);
3507 /// \brief Register a store.
3508 void addStore(Value *Ptr) {
3509 Accesses.insert(MemAccessInfo(Ptr, true));
3512 /// \brief Check whether we can check the pointers at runtime for
3513 /// non-intersection.
3514 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3515 unsigned &NumComparisons, ScalarEvolution *SE,
3516 Loop *TheLoop, ValueToValueMap &Strides,
3517 bool ShouldCheckStride = false);
3519 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3520 /// and builds sets of dependent accesses.
3521 void buildDependenceSets() {
3522 // Process read-write pointers first.
3523 processMemAccesses(false);
3524 // Next, process read pointers.
3525 processMemAccesses(true);
3528 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3530 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3531 void resetDepChecks() { CheckDeps.clear(); }
3533 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3536 typedef SetVector<MemAccessInfo> PtrAccessSet;
3537 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3539 /// \brief Go over all memory access or only the deferred ones if
3540 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3541 /// and build sets of dependency check candidates.
3542 void processMemAccesses(bool UseDeferred);
3544 /// Set of all accesses.
3545 PtrAccessSet Accesses;
3547 /// Set of access to check after all writes have been processed.
3548 PtrAccessSet DeferredAccesses;
3550 /// Map of pointers to last access encountered.
3551 UnderlyingObjToAccessMap ObjToLastAccess;
3553 /// Set of accesses that need a further dependence check.
3554 MemAccessInfoSet CheckDeps;
3556 /// Set of pointers that are read only.
3557 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3559 /// Set of underlying objects already written to.
3560 SmallPtrSet<Value*, 16> WriteObjects;
3564 /// Sets of potentially dependent accesses - members of one set share an
3565 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3566 /// dependence check.
3567 DepCandidates &DepCands;
3569 bool AreAllWritesIdentified;
3570 bool AreAllReadsIdentified;
3571 bool IsRTCheckNeeded;
3574 } // end anonymous namespace
3576 /// \brief Check whether a pointer can participate in a runtime bounds check.
3577 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3579 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3580 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3584 return AR->isAffine();
3587 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3588 /// the address space.
3589 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3590 const Loop *Lp, ValueToValueMap &StridesMap);
3592 bool AccessAnalysis::canCheckPtrAtRT(
3593 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3594 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3595 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3596 // Find pointers with computable bounds. We are going to use this information
3597 // to place a runtime bound check.
3598 unsigned NumReadPtrChecks = 0;
3599 unsigned NumWritePtrChecks = 0;
3600 bool CanDoRT = true;
3602 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3603 // We assign consecutive id to access from different dependence sets.
3604 // Accesses within the same set don't need a runtime check.
3605 unsigned RunningDepId = 1;
3606 DenseMap<Value *, unsigned> DepSetId;
3608 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3610 const MemAccessInfo &Access = *AI;
3611 Value *Ptr = Access.getPointer();
3612 bool IsWrite = Access.getInt();
3614 // Just add write checks if we have both.
3615 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3619 ++NumWritePtrChecks;
3623 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3624 // When we run after a failing dependency check we have to make sure we
3625 // don't have wrapping pointers.
3626 (!ShouldCheckStride ||
3627 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3628 // The id of the dependence set.
3631 if (IsDepCheckNeeded) {
3632 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3633 unsigned &LeaderId = DepSetId[Leader];
3635 LeaderId = RunningDepId++;
3638 // Each access has its own dependence set.
3639 DepId = RunningDepId++;
3641 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3643 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3649 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3650 NumComparisons = 0; // Only one dependence set.
3652 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3653 NumWritePtrChecks - 1));
3656 // If the pointers that we would use for the bounds comparison have different
3657 // address spaces, assume the values aren't directly comparable, so we can't
3658 // use them for the runtime check. We also have to assume they could
3659 // overlap. In the future there should be metadata for whether address spaces
3661 unsigned NumPointers = RtCheck.Pointers.size();
3662 for (unsigned i = 0; i < NumPointers; ++i) {
3663 for (unsigned j = i + 1; j < NumPointers; ++j) {
3664 // Only need to check pointers between two different dependency sets.
3665 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3668 Value *PtrI = RtCheck.Pointers[i];
3669 Value *PtrJ = RtCheck.Pointers[j];
3671 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3672 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3674 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3675 " different address spaces\n");
3684 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3685 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3688 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3689 // We process the set twice: first we process read-write pointers, last we
3690 // process read-only pointers. This allows us to skip dependence tests for
3691 // read-only pointers.
3693 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3694 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3695 const MemAccessInfo &Access = *AI;
3696 Value *Ptr = Access.getPointer();
3697 bool IsWrite = Access.getInt();
3699 DepCands.insert(Access);
3701 // Memorize read-only pointers for later processing and skip them in the
3702 // first round (they need to be checked after we have seen all write
3703 // pointers). Note: we also mark pointer that are not consecutive as
3704 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3705 // second check for "!IsWrite".
3706 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3707 if (!UseDeferred && IsReadOnlyPtr) {
3708 DeferredAccesses.insert(Access);
3712 bool NeedDepCheck = false;
3713 // Check whether there is the possiblity of dependency because of underlying
3714 // objects being the same.
3715 typedef SmallVector<Value*, 16> ValueVector;
3716 ValueVector TempObjects;
3717 GetUnderlyingObjects(Ptr, TempObjects, DL);
3718 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3720 Value *UnderlyingObj = *UI;
3722 // If this is a write then it needs to be an identified object. If this a
3723 // read and all writes (so far) are identified function scope objects we
3724 // don't need an identified underlying object but only an Argument (the
3725 // next write is going to invalidate this assumption if it is
3727 // This is a micro-optimization for the case where all writes are
3728 // identified and we have one argument pointer.
3729 // Otherwise, we do need a runtime check.
3730 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3731 (!IsWrite && (!AreAllWritesIdentified ||
3732 !isa<Argument>(UnderlyingObj)) &&
3733 !isIdentifiedObject(UnderlyingObj))) {
3734 DEBUG(dbgs() << "LV: Found an unidentified " <<
3735 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3737 IsRTCheckNeeded = (IsRTCheckNeeded ||
3738 !isIdentifiedObject(UnderlyingObj) ||
3739 !AreAllReadsIdentified);
3742 AreAllWritesIdentified = false;
3744 AreAllReadsIdentified = false;
3747 // If this is a write - check other reads and writes for conflicts. If
3748 // this is a read only check other writes for conflicts (but only if there
3749 // is no other write to the ptr - this is an optimization to catch "a[i] =
3750 // a[i] + " without having to do a dependence check).
3751 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3752 NeedDepCheck = true;
3755 WriteObjects.insert(UnderlyingObj);
3757 // Create sets of pointers connected by shared underlying objects.
3758 UnderlyingObjToAccessMap::iterator Prev =
3759 ObjToLastAccess.find(UnderlyingObj);
3760 if (Prev != ObjToLastAccess.end())
3761 DepCands.unionSets(Access, Prev->second);
3763 ObjToLastAccess[UnderlyingObj] = Access;
3767 CheckDeps.insert(Access);
3772 /// \brief Checks memory dependences among accesses to the same underlying
3773 /// object to determine whether there vectorization is legal or not (and at
3774 /// which vectorization factor).
3776 /// This class works under the assumption that we already checked that memory
3777 /// locations with different underlying pointers are "must-not alias".
3778 /// We use the ScalarEvolution framework to symbolically evalutate access
3779 /// functions pairs. Since we currently don't restructure the loop we can rely
3780 /// on the program order of memory accesses to determine their safety.
3781 /// At the moment we will only deem accesses as safe for:
3782 /// * A negative constant distance assuming program order.
3784 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3785 /// a[i] = tmp; y = a[i];
3787 /// The latter case is safe because later checks guarantuee that there can't
3788 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3789 /// the same variable: a header phi can only be an induction or a reduction, a
3790 /// reduction can't have a memory sink, an induction can't have a memory
3791 /// source). This is important and must not be violated (or we have to
3792 /// resort to checking for cycles through memory).
3794 /// * A positive constant distance assuming program order that is bigger
3795 /// than the biggest memory access.
3797 /// tmp = a[i] OR b[i] = x
3798 /// a[i+2] = tmp y = b[i+2];
3800 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3802 /// * Zero distances and all accesses have the same size.
3804 class MemoryDepChecker {
3806 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3807 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3809 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3810 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3811 ShouldRetryWithRuntimeCheck(false) {}
3813 /// \brief Register the location (instructions are given increasing numbers)
3814 /// of a write access.
3815 void addAccess(StoreInst *SI) {
3816 Value *Ptr = SI->getPointerOperand();
3817 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3818 InstMap.push_back(SI);
3822 /// \brief Register the location (instructions are given increasing numbers)
3823 /// of a write access.
3824 void addAccess(LoadInst *LI) {
3825 Value *Ptr = LI->getPointerOperand();
3826 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3827 InstMap.push_back(LI);
3831 /// \brief Check whether the dependencies between the accesses are safe.
3833 /// Only checks sets with elements in \p CheckDeps.
3834 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3835 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
3837 /// \brief The maximum number of bytes of a vector register we can vectorize
3838 /// the accesses safely with.
3839 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3841 /// \brief In same cases when the dependency check fails we can still
3842 /// vectorize the loop with a dynamic array access check.
3843 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3846 ScalarEvolution *SE;
3848 const Loop *InnermostLoop;
3850 /// \brief Maps access locations (ptr, read/write) to program order.
3851 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3853 /// \brief Memory access instructions in program order.
3854 SmallVector<Instruction *, 16> InstMap;
3856 /// \brief The program order index to be used for the next instruction.
3859 // We can access this many bytes in parallel safely.
3860 unsigned MaxSafeDepDistBytes;
3862 /// \brief If we see a non-constant dependence distance we can still try to
3863 /// vectorize this loop with runtime checks.
3864 bool ShouldRetryWithRuntimeCheck;
3866 /// \brief Check whether there is a plausible dependence between the two
3869 /// Access \p A must happen before \p B in program order. The two indices
3870 /// identify the index into the program order map.
3872 /// This function checks whether there is a plausible dependence (or the
3873 /// absence of such can't be proved) between the two accesses. If there is a
3874 /// plausible dependence but the dependence distance is bigger than one
3875 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3876 /// distance is smaller than any other distance encountered so far).
3877 /// Otherwise, this function returns true signaling a possible dependence.
3878 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3879 const MemAccessInfo &B, unsigned BIdx,
3880 ValueToValueMap &Strides);
3882 /// \brief Check whether the data dependence could prevent store-load
3884 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3887 } // end anonymous namespace
3889 static bool isInBoundsGep(Value *Ptr) {
3890 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3891 return GEP->isInBounds();
3895 /// \brief Check whether the access through \p Ptr has a constant stride.
3896 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3897 const Loop *Lp, ValueToValueMap &StridesMap) {
3898 const Type *Ty = Ptr->getType();
3899 assert(Ty->isPointerTy() && "Unexpected non-ptr");
3901 // Make sure that the pointer does not point to aggregate types.
3902 const PointerType *PtrTy = cast<PointerType>(Ty);
3903 if (PtrTy->getElementType()->isAggregateType()) {
3904 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3909 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
3911 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3913 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3914 << *Ptr << " SCEV: " << *PtrScev << "\n");
3918 // The accesss function must stride over the innermost loop.
3919 if (Lp != AR->getLoop()) {
3920 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3921 *Ptr << " SCEV: " << *PtrScev << "\n");
3924 // The address calculation must not wrap. Otherwise, a dependence could be
3926 // An inbounds getelementptr that is a AddRec with a unit stride
3927 // cannot wrap per definition. The unit stride requirement is checked later.
3928 // An getelementptr without an inbounds attribute and unit stride would have
3929 // to access the pointer value "0" which is undefined behavior in address
3930 // space 0, therefore we can also vectorize this case.
3931 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3932 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3933 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3934 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3935 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3936 << *Ptr << " SCEV: " << *PtrScev << "\n");
3940 // Check the step is constant.
3941 const SCEV *Step = AR->getStepRecurrence(*SE);
3943 // Calculate the pointer stride and check if it is consecutive.
3944 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3946 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3947 " SCEV: " << *PtrScev << "\n");
3951 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3952 const APInt &APStepVal = C->getValue()->getValue();
3954 // Huge step value - give up.
3955 if (APStepVal.getBitWidth() > 64)
3958 int64_t StepVal = APStepVal.getSExtValue();
3961 int64_t Stride = StepVal / Size;
3962 int64_t Rem = StepVal % Size;
3966 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3967 // know we can't "wrap around the address space". In case of address space
3968 // zero we know that this won't happen without triggering undefined behavior.
3969 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3970 Stride != 1 && Stride != -1)
3976 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3977 unsigned TypeByteSize) {
3978 // If loads occur at a distance that is not a multiple of a feasible vector
3979 // factor store-load forwarding does not take place.
3980 // Positive dependences might cause troubles because vectorizing them might
3981 // prevent store-load forwarding making vectorized code run a lot slower.
3982 // a[i] = a[i-3] ^ a[i-8];
3983 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3984 // hence on your typical architecture store-load forwarding does not take
3985 // place. Vectorizing in such cases does not make sense.
3986 // Store-load forwarding distance.
3987 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3988 // Maximum vector factor.
3989 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3990 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3991 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3993 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3995 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3996 MaxVFWithoutSLForwardIssues = (vf >>=1);
4001 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4002 DEBUG(dbgs() << "LV: Distance " << Distance <<
4003 " that could cause a store-load forwarding conflict\n");
4007 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4008 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4009 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4013 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4014 const MemAccessInfo &B, unsigned BIdx,
4015 ValueToValueMap &Strides) {
4016 assert (AIdx < BIdx && "Must pass arguments in program order");
4018 Value *APtr = A.getPointer();
4019 Value *BPtr = B.getPointer();
4020 bool AIsWrite = A.getInt();
4021 bool BIsWrite = B.getInt();
4023 // Two reads are independent.
4024 if (!AIsWrite && !BIsWrite)
4027 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4028 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4030 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4031 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4033 const SCEV *Src = AScev;
4034 const SCEV *Sink = BScev;
4036 // If the induction step is negative we have to invert source and sink of the
4038 if (StrideAPtr < 0) {
4041 std::swap(APtr, BPtr);
4042 std::swap(Src, Sink);
4043 std::swap(AIsWrite, BIsWrite);
4044 std::swap(AIdx, BIdx);
4045 std::swap(StrideAPtr, StrideBPtr);
4048 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4050 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4051 << "(Induction step: " << StrideAPtr << ")\n");
4052 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4053 << *InstMap[BIdx] << ": " << *Dist << "\n");
4055 // Need consecutive accesses. We don't want to vectorize
4056 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4057 // the address space.
4058 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4059 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4063 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4065 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4066 ShouldRetryWithRuntimeCheck = true;
4070 Type *ATy = APtr->getType()->getPointerElementType();
4071 Type *BTy = BPtr->getType()->getPointerElementType();
4072 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4074 // Negative distances are not plausible dependencies.
4075 const APInt &Val = C->getValue()->getValue();
4076 if (Val.isNegative()) {
4077 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4078 if (IsTrueDataDependence &&
4079 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4083 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4087 // Write to the same location with the same size.
4088 // Could be improved to assert type sizes are the same (i32 == float, etc).
4092 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4096 assert(Val.isStrictlyPositive() && "Expect a positive value");
4098 // Positive distance bigger than max vectorization factor.
4101 "LV: ReadWrite-Write positive dependency with different types\n");
4105 unsigned Distance = (unsigned) Val.getZExtValue();
4107 // Bail out early if passed-in parameters make vectorization not feasible.
4108 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4109 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4111 // The distance must be bigger than the size needed for a vectorized version
4112 // of the operation and the size of the vectorized operation must not be
4113 // bigger than the currrent maximum size.
4114 if (Distance < 2*TypeByteSize ||
4115 2*TypeByteSize > MaxSafeDepDistBytes ||
4116 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4117 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4118 << Val.getSExtValue() << '\n');
4122 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4123 Distance : MaxSafeDepDistBytes;
4125 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4126 if (IsTrueDataDependence &&
4127 couldPreventStoreLoadForward(Distance, TypeByteSize))
4130 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4131 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4136 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4137 MemAccessInfoSet &CheckDeps,
4138 ValueToValueMap &Strides) {
4140 MaxSafeDepDistBytes = -1U;
4141 while (!CheckDeps.empty()) {
4142 MemAccessInfo CurAccess = *CheckDeps.begin();
4144 // Get the relevant memory access set.
4145 EquivalenceClasses<MemAccessInfo>::iterator I =
4146 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4148 // Check accesses within this set.
4149 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4150 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4152 // Check every access pair.
4154 CheckDeps.erase(*AI);
4155 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
4157 // Check every accessing instruction pair in program order.
4158 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4159 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4160 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4161 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4162 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4164 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4175 bool LoopVectorizationLegality::canVectorizeMemory() {
4177 typedef SmallVector<Value*, 16> ValueVector;
4178 typedef SmallPtrSet<Value*, 16> ValueSet;
4180 // Holds the Load and Store *instructions*.
4184 // Holds all the different accesses in the loop.
4185 unsigned NumReads = 0;
4186 unsigned NumReadWrites = 0;
4188 PtrRtCheck.Pointers.clear();
4189 PtrRtCheck.Need = false;
4191 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4192 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4195 for (Loop::block_iterator bb = TheLoop->block_begin(),
4196 be = TheLoop->block_end(); bb != be; ++bb) {
4198 // Scan the BB and collect legal loads and stores.
4199 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4202 // If this is a load, save it. If this instruction can read from memory
4203 // but is not a load, then we quit. Notice that we don't handle function
4204 // calls that read or write.
4205 if (it->mayReadFromMemory()) {
4206 // Many math library functions read the rounding mode. We will only
4207 // vectorize a loop if it contains known function calls that don't set
4208 // the flag. Therefore, it is safe to ignore this read from memory.
4209 CallInst *Call = dyn_cast<CallInst>(it);
4210 if (Call && getIntrinsicIDForCall(Call, TLI))
4213 LoadInst *Ld = dyn_cast<LoadInst>(it);
4214 if (!Ld) return false;
4215 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4216 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4219 Loads.push_back(Ld);
4220 DepChecker.addAccess(Ld);
4224 // Save 'store' instructions. Abort if other instructions write to memory.
4225 if (it->mayWriteToMemory()) {
4226 StoreInst *St = dyn_cast<StoreInst>(it);
4227 if (!St) return false;
4228 if (!St->isSimple() && !IsAnnotatedParallel) {
4229 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4232 Stores.push_back(St);
4233 DepChecker.addAccess(St);
4238 // Now we have two lists that hold the loads and the stores.
4239 // Next, we find the pointers that they use.
4241 // Check if we see any stores. If there are no stores, then we don't
4242 // care if the pointers are *restrict*.
4243 if (!Stores.size()) {
4244 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4248 AccessAnalysis::DepCandidates DependentAccesses;
4249 AccessAnalysis Accesses(DL, DependentAccesses);
4251 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4252 // multiple times on the same object. If the ptr is accessed twice, once
4253 // for read and once for write, it will only appear once (on the write
4254 // list). This is okay, since we are going to check for conflicts between
4255 // writes and between reads and writes, but not between reads and reads.
4258 ValueVector::iterator I, IE;
4259 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4260 StoreInst *ST = cast<StoreInst>(*I);
4261 Value* Ptr = ST->getPointerOperand();
4263 if (isUniform(Ptr)) {
4264 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4268 // If we did *not* see this pointer before, insert it to the read-write
4269 // list. At this phase it is only a 'write' list.
4270 if (Seen.insert(Ptr)) {
4272 Accesses.addStore(Ptr);
4276 if (IsAnnotatedParallel) {
4278 << "LV: A loop annotated parallel, ignore memory dependency "
4283 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4284 LoadInst *LD = cast<LoadInst>(*I);
4285 Value* Ptr = LD->getPointerOperand();
4286 // If we did *not* see this pointer before, insert it to the
4287 // read list. If we *did* see it before, then it is already in
4288 // the read-write list. This allows us to vectorize expressions
4289 // such as A[i] += x; Because the address of A[i] is a read-write
4290 // pointer. This only works if the index of A[i] is consecutive.
4291 // If the address of i is unknown (for example A[B[i]]) then we may
4292 // read a few words, modify, and write a few words, and some of the
4293 // words may be written to the same address.
4294 bool IsReadOnlyPtr = false;
4295 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4297 IsReadOnlyPtr = true;
4299 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4302 // If we write (or read-write) to a single destination and there are no
4303 // other reads in this loop then is it safe to vectorize.
4304 if (NumReadWrites == 1 && NumReads == 0) {
4305 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4309 // Build dependence sets and check whether we need a runtime pointer bounds
4311 Accesses.buildDependenceSets();
4312 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4314 // Find pointers with computable bounds. We are going to use this information
4315 // to place a runtime bound check.
4316 unsigned NumComparisons = 0;
4317 bool CanDoRT = false;
4319 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4322 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4323 " pointer comparisons.\n");
4325 // If we only have one set of dependences to check pointers among we don't
4326 // need a runtime check.
4327 if (NumComparisons == 0 && NeedRTCheck)
4328 NeedRTCheck = false;
4330 // Check that we did not collect too many pointers or found an unsizeable
4332 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4338 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4341 if (NeedRTCheck && !CanDoRT) {
4342 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4343 "the array bounds.\n");
4348 PtrRtCheck.Need = NeedRTCheck;
4350 bool CanVecMem = true;
4351 if (Accesses.isDependencyCheckNeeded()) {
4352 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4353 CanVecMem = DepChecker.areDepsSafe(
4354 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4355 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4357 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4358 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4361 // Clear the dependency checks. We assume they are not needed.
4362 Accesses.resetDepChecks();
4365 PtrRtCheck.Need = true;
4367 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4368 TheLoop, Strides, true);
4369 // Check that we did not collect too many pointers or found an unsizeable
4371 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4372 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4381 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4382 " need a runtime memory check.\n");
4387 static bool hasMultipleUsesOf(Instruction *I,
4388 SmallPtrSet<Instruction *, 8> &Insts) {
4389 unsigned NumUses = 0;
4390 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4391 if (Insts.count(dyn_cast<Instruction>(*Use)))
4400 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4401 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4402 if (!Set.count(dyn_cast<Instruction>(*Use)))
4407 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4408 ReductionKind Kind) {
4409 if (Phi->getNumIncomingValues() != 2)
4412 // Reduction variables are only found in the loop header block.
4413 if (Phi->getParent() != TheLoop->getHeader())
4416 // Obtain the reduction start value from the value that comes from the loop
4418 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4420 // ExitInstruction is the single value which is used outside the loop.
4421 // We only allow for a single reduction value to be used outside the loop.
4422 // This includes users of the reduction, variables (which form a cycle
4423 // which ends in the phi node).
4424 Instruction *ExitInstruction = 0;
4425 // Indicates that we found a reduction operation in our scan.
4426 bool FoundReduxOp = false;
4428 // We start with the PHI node and scan for all of the users of this
4429 // instruction. All users must be instructions that can be used as reduction
4430 // variables (such as ADD). We must have a single out-of-block user. The cycle
4431 // must include the original PHI.
4432 bool FoundStartPHI = false;
4434 // To recognize min/max patterns formed by a icmp select sequence, we store
4435 // the number of instruction we saw from the recognized min/max pattern,
4436 // to make sure we only see exactly the two instructions.
4437 unsigned NumCmpSelectPatternInst = 0;
4438 ReductionInstDesc ReduxDesc(false, 0);
4440 SmallPtrSet<Instruction *, 8> VisitedInsts;
4441 SmallVector<Instruction *, 8> Worklist;
4442 Worklist.push_back(Phi);
4443 VisitedInsts.insert(Phi);
4445 // A value in the reduction can be used:
4446 // - By the reduction:
4447 // - Reduction operation:
4448 // - One use of reduction value (safe).
4449 // - Multiple use of reduction value (not safe).
4451 // - All uses of the PHI must be the reduction (safe).
4452 // - Otherwise, not safe.
4453 // - By one instruction outside of the loop (safe).
4454 // - By further instructions outside of the loop (not safe).
4455 // - By an instruction that is not part of the reduction (not safe).
4457 // * An instruction type other than PHI or the reduction operation.
4458 // * A PHI in the header other than the initial PHI.
4459 while (!Worklist.empty()) {
4460 Instruction *Cur = Worklist.back();
4461 Worklist.pop_back();
4464 // If the instruction has no users then this is a broken chain and can't be
4465 // a reduction variable.
4466 if (Cur->use_empty())
4469 bool IsAPhi = isa<PHINode>(Cur);
4471 // A header PHI use other than the original PHI.
4472 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4475 // Reductions of instructions such as Div, and Sub is only possible if the
4476 // LHS is the reduction variable.
4477 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4478 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4479 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4482 // Any reduction instruction must be of one of the allowed kinds.
4483 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4484 if (!ReduxDesc.IsReduction)
4487 // A reduction operation must only have one use of the reduction value.
4488 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4489 hasMultipleUsesOf(Cur, VisitedInsts))
4492 // All inputs to a PHI node must be a reduction value.
4493 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4496 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4497 isa<SelectInst>(Cur)))
4498 ++NumCmpSelectPatternInst;
4499 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4500 isa<SelectInst>(Cur)))
4501 ++NumCmpSelectPatternInst;
4503 // Check whether we found a reduction operator.
4504 FoundReduxOp |= !IsAPhi;
4506 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4507 // onto the stack. This way we are going to have seen all inputs to PHI
4508 // nodes once we get to them.
4509 SmallVector<Instruction *, 8> NonPHIs;
4510 SmallVector<Instruction *, 8> PHIs;
4511 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4513 Instruction *Usr = cast<Instruction>(*UI);
4515 // Check if we found the exit user.
4516 BasicBlock *Parent = Usr->getParent();
4517 if (!TheLoop->contains(Parent)) {
4518 // Exit if you find multiple outside users or if the header phi node is
4519 // being used. In this case the user uses the value of the previous
4520 // iteration, in which case we would loose "VF-1" iterations of the
4521 // reduction operation if we vectorize.
4522 if (ExitInstruction != 0 || Cur == Phi)
4525 // The instruction used by an outside user must be the last instruction
4526 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4527 // operations on the value.
4528 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4531 ExitInstruction = Cur;
4535 // Process instructions only once (termination).
4536 if (VisitedInsts.insert(Usr)) {
4537 if (isa<PHINode>(Usr))
4538 PHIs.push_back(Usr);
4540 NonPHIs.push_back(Usr);
4542 // Remember that we completed the cycle.
4544 FoundStartPHI = true;
4546 Worklist.append(PHIs.begin(), PHIs.end());
4547 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4550 // This means we have seen one but not the other instruction of the
4551 // pattern or more than just a select and cmp.
4552 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4553 NumCmpSelectPatternInst != 2)
4556 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4559 // We found a reduction var if we have reached the original phi node and we
4560 // only have a single instruction with out-of-loop users.
4562 // This instruction is allowed to have out-of-loop users.
4563 AllowedExit.insert(ExitInstruction);
4565 // Save the description of this reduction variable.
4566 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4567 ReduxDesc.MinMaxKind);
4568 Reductions[Phi] = RD;
4569 // We've ended the cycle. This is a reduction variable if we have an
4570 // outside user and it has a binary op.
4575 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4576 /// pattern corresponding to a min(X, Y) or max(X, Y).
4577 LoopVectorizationLegality::ReductionInstDesc
4578 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4579 ReductionInstDesc &Prev) {
4581 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4582 "Expect a select instruction");
4583 Instruction *Cmp = 0;
4584 SelectInst *Select = 0;
4586 // We must handle the select(cmp()) as a single instruction. Advance to the
4588 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4589 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4590 return ReductionInstDesc(false, I);
4591 return ReductionInstDesc(Select, Prev.MinMaxKind);
4594 // Only handle single use cases for now.
4595 if (!(Select = dyn_cast<SelectInst>(I)))
4596 return ReductionInstDesc(false, I);
4597 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4598 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4599 return ReductionInstDesc(false, I);
4600 if (!Cmp->hasOneUse())
4601 return ReductionInstDesc(false, I);
4606 // Look for a min/max pattern.
4607 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4608 return ReductionInstDesc(Select, MRK_UIntMin);
4609 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4610 return ReductionInstDesc(Select, MRK_UIntMax);
4611 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4612 return ReductionInstDesc(Select, MRK_SIntMax);
4613 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4614 return ReductionInstDesc(Select, MRK_SIntMin);
4615 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4616 return ReductionInstDesc(Select, MRK_FloatMin);
4617 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4618 return ReductionInstDesc(Select, MRK_FloatMax);
4619 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4620 return ReductionInstDesc(Select, MRK_FloatMin);
4621 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4622 return ReductionInstDesc(Select, MRK_FloatMax);
4624 return ReductionInstDesc(false, I);
4627 LoopVectorizationLegality::ReductionInstDesc
4628 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4630 ReductionInstDesc &Prev) {
4631 bool FP = I->getType()->isFloatingPointTy();
4632 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4633 switch (I->getOpcode()) {
4635 return ReductionInstDesc(false, I);
4636 case Instruction::PHI:
4637 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4638 Kind != RK_FloatMinMax))
4639 return ReductionInstDesc(false, I);
4640 return ReductionInstDesc(I, Prev.MinMaxKind);
4641 case Instruction::Sub:
4642 case Instruction::Add:
4643 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4644 case Instruction::Mul:
4645 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4646 case Instruction::And:
4647 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4648 case Instruction::Or:
4649 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4650 case Instruction::Xor:
4651 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4652 case Instruction::FMul:
4653 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4654 case Instruction::FAdd:
4655 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4656 case Instruction::FCmp:
4657 case Instruction::ICmp:
4658 case Instruction::Select:
4659 if (Kind != RK_IntegerMinMax &&
4660 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4661 return ReductionInstDesc(false, I);
4662 return isMinMaxSelectCmpPattern(I, Prev);
4666 LoopVectorizationLegality::InductionKind
4667 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4668 Type *PhiTy = Phi->getType();
4669 // We only handle integer and pointer inductions variables.
4670 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4671 return IK_NoInduction;
4673 // Check that the PHI is consecutive.
4674 const SCEV *PhiScev = SE->getSCEV(Phi);
4675 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4677 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4678 return IK_NoInduction;
4680 const SCEV *Step = AR->getStepRecurrence(*SE);
4682 // Integer inductions need to have a stride of one.
4683 if (PhiTy->isIntegerTy()) {
4685 return IK_IntInduction;
4686 if (Step->isAllOnesValue())
4687 return IK_ReverseIntInduction;
4688 return IK_NoInduction;
4691 // Calculate the pointer stride and check if it is consecutive.
4692 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4694 return IK_NoInduction;
4696 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4697 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4698 if (C->getValue()->equalsInt(Size))
4699 return IK_PtrInduction;
4700 else if (C->getValue()->equalsInt(0 - Size))
4701 return IK_ReversePtrInduction;
4703 return IK_NoInduction;
4706 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4707 Value *In0 = const_cast<Value*>(V);
4708 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4712 return Inductions.count(PN);
4715 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4716 assert(TheLoop->contains(BB) && "Unknown block used");
4718 // Blocks that do not dominate the latch need predication.
4719 BasicBlock* Latch = TheLoop->getLoopLatch();
4720 return !DT->dominates(BB, Latch);
4723 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4724 SmallPtrSet<Value *, 8>& SafePtrs) {
4725 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4726 // We might be able to hoist the load.
4727 if (it->mayReadFromMemory()) {
4728 LoadInst *LI = dyn_cast<LoadInst>(it);
4729 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4733 // We don't predicate stores at the moment.
4734 if (it->mayWriteToMemory() || it->mayThrow())
4737 // Check that we don't have a constant expression that can trap as operand.
4738 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4740 if (Constant *C = dyn_cast<Constant>(*OI))
4745 // The instructions below can trap.
4746 switch (it->getOpcode()) {
4748 case Instruction::UDiv:
4749 case Instruction::SDiv:
4750 case Instruction::URem:
4751 case Instruction::SRem:
4759 LoopVectorizationCostModel::VectorizationFactor
4760 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4762 // Width 1 means no vectorize
4763 VectorizationFactor Factor = { 1U, 0U };
4764 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4765 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4769 // Find the trip count.
4770 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4771 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4773 unsigned WidestType = getWidestType();
4774 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4775 unsigned MaxSafeDepDist = -1U;
4776 if (Legal->getMaxSafeDepDistBytes() != -1U)
4777 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4778 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4779 WidestRegister : MaxSafeDepDist);
4780 unsigned MaxVectorSize = WidestRegister / WidestType;
4781 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4782 DEBUG(dbgs() << "LV: The Widest register is: "
4783 << WidestRegister << " bits.\n");
4785 if (MaxVectorSize == 0) {
4786 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4790 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4791 " into one vector!");
4793 unsigned VF = MaxVectorSize;
4795 // If we optimize the program for size, avoid creating the tail loop.
4797 // If we are unable to calculate the trip count then don't try to vectorize.
4799 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4803 // Find the maximum SIMD width that can fit within the trip count.
4804 VF = TC % MaxVectorSize;
4809 // If the trip count that we found modulo the vectorization factor is not
4810 // zero then we require a tail.
4812 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4818 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4819 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4821 Factor.Width = UserVF;
4825 float Cost = expectedCost(1);
4827 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
4828 for (unsigned i=2; i <= VF; i*=2) {
4829 // Notice that the vector loop needs to be executed less times, so
4830 // we need to divide the cost of the vector loops by the width of
4831 // the vector elements.
4832 float VectorCost = expectedCost(i) / (float)i;
4833 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4834 (int)VectorCost << ".\n");
4835 if (VectorCost < Cost) {
4841 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4842 Factor.Width = Width;
4843 Factor.Cost = Width * Cost;
4847 unsigned LoopVectorizationCostModel::getWidestType() {
4848 unsigned MaxWidth = 8;
4851 for (Loop::block_iterator bb = TheLoop->block_begin(),
4852 be = TheLoop->block_end(); bb != be; ++bb) {
4853 BasicBlock *BB = *bb;
4855 // For each instruction in the loop.
4856 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4857 Type *T = it->getType();
4859 // Only examine Loads, Stores and PHINodes.
4860 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4863 // Examine PHI nodes that are reduction variables.
4864 if (PHINode *PN = dyn_cast<PHINode>(it))
4865 if (!Legal->getReductionVars()->count(PN))
4868 // Examine the stored values.
4869 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4870 T = ST->getValueOperand()->getType();
4872 // Ignore loaded pointer types and stored pointer types that are not
4873 // consecutive. However, we do want to take consecutive stores/loads of
4874 // pointer vectors into account.
4875 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4878 MaxWidth = std::max(MaxWidth,
4879 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4887 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4890 unsigned LoopCost) {
4892 // -- The unroll heuristics --
4893 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4894 // There are many micro-architectural considerations that we can't predict
4895 // at this level. For example frontend pressure (on decode or fetch) due to
4896 // code size, or the number and capabilities of the execution ports.
4898 // We use the following heuristics to select the unroll factor:
4899 // 1. If the code has reductions the we unroll in order to break the cross
4900 // iteration dependency.
4901 // 2. If the loop is really small then we unroll in order to reduce the loop
4903 // 3. We don't unroll if we think that we will spill registers to memory due
4904 // to the increased register pressure.
4906 // Use the user preference, unless 'auto' is selected.
4910 // When we optimize for size we don't unroll.
4914 // We used the distance for the unroll factor.
4915 if (Legal->getMaxSafeDepDistBytes() != -1U)
4918 // Do not unroll loops with a relatively small trip count.
4919 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4920 TheLoop->getLoopLatch());
4921 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4924 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4925 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4926 " vector registers\n");
4928 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4929 // We divide by these constants so assume that we have at least one
4930 // instruction that uses at least one register.
4931 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4932 R.NumInstructions = std::max(R.NumInstructions, 1U);
4934 // We calculate the unroll factor using the following formula.
4935 // Subtract the number of loop invariants from the number of available
4936 // registers. These registers are used by all of the unrolled instances.
4937 // Next, divide the remaining registers by the number of registers that is
4938 // required by the loop, in order to estimate how many parallel instances
4939 // fit without causing spills.
4940 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4942 // Clamp the unroll factor ranges to reasonable factors.
4943 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4945 // If we did not calculate the cost for VF (because the user selected the VF)
4946 // then we calculate the cost of VF here.
4948 LoopCost = expectedCost(VF);
4950 // Clamp the calculated UF to be between the 1 and the max unroll factor
4951 // that the target allows.
4952 if (UF > MaxUnrollSize)
4957 bool HasReductions = Legal->getReductionVars()->size();
4959 // Decide if we want to unroll if we decided that it is legal to vectorize
4960 // but not profitable.
4962 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4963 LoopCost > SmallLoopCost)
4969 if (HasReductions) {
4970 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4974 // We want to unroll tiny loops in order to reduce the loop overhead.
4975 // We assume that the cost overhead is 1 and we use the cost model
4976 // to estimate the cost of the loop and unroll until the cost of the
4977 // loop overhead is about 5% of the cost of the loop.
4978 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4979 if (LoopCost < SmallLoopCost) {
4980 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4981 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4982 return std::min(NewUF, UF);
4985 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4989 LoopVectorizationCostModel::RegisterUsage
4990 LoopVectorizationCostModel::calculateRegisterUsage() {
4991 // This function calculates the register usage by measuring the highest number
4992 // of values that are alive at a single location. Obviously, this is a very
4993 // rough estimation. We scan the loop in a topological order in order and
4994 // assign a number to each instruction. We use RPO to ensure that defs are
4995 // met before their users. We assume that each instruction that has in-loop
4996 // users starts an interval. We record every time that an in-loop value is
4997 // used, so we have a list of the first and last occurrences of each
4998 // instruction. Next, we transpose this data structure into a multi map that
4999 // holds the list of intervals that *end* at a specific location. This multi
5000 // map allows us to perform a linear search. We scan the instructions linearly
5001 // and record each time that a new interval starts, by placing it in a set.
5002 // If we find this value in the multi-map then we remove it from the set.
5003 // The max register usage is the maximum size of the set.
5004 // We also search for instructions that are defined outside the loop, but are
5005 // used inside the loop. We need this number separately from the max-interval
5006 // usage number because when we unroll, loop-invariant values do not take
5008 LoopBlocksDFS DFS(TheLoop);
5012 R.NumInstructions = 0;
5014 // Each 'key' in the map opens a new interval. The values
5015 // of the map are the index of the 'last seen' usage of the
5016 // instruction that is the key.
5017 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5018 // Maps instruction to its index.
5019 DenseMap<unsigned, Instruction*> IdxToInstr;
5020 // Marks the end of each interval.
5021 IntervalMap EndPoint;
5022 // Saves the list of instruction indices that are used in the loop.
5023 SmallSet<Instruction*, 8> Ends;
5024 // Saves the list of values that are used in the loop but are
5025 // defined outside the loop, such as arguments and constants.
5026 SmallPtrSet<Value*, 8> LoopInvariants;
5029 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5030 be = DFS.endRPO(); bb != be; ++bb) {
5031 R.NumInstructions += (*bb)->size();
5032 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5034 Instruction *I = it;
5035 IdxToInstr[Index++] = I;
5037 // Save the end location of each USE.
5038 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5039 Value *U = I->getOperand(i);
5040 Instruction *Instr = dyn_cast<Instruction>(U);
5042 // Ignore non-instruction values such as arguments, constants, etc.
5043 if (!Instr) continue;
5045 // If this instruction is outside the loop then record it and continue.
5046 if (!TheLoop->contains(Instr)) {
5047 LoopInvariants.insert(Instr);
5051 // Overwrite previous end points.
5052 EndPoint[Instr] = Index;
5058 // Saves the list of intervals that end with the index in 'key'.
5059 typedef SmallVector<Instruction*, 2> InstrList;
5060 DenseMap<unsigned, InstrList> TransposeEnds;
5062 // Transpose the EndPoints to a list of values that end at each index.
5063 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5065 TransposeEnds[it->second].push_back(it->first);
5067 SmallSet<Instruction*, 8> OpenIntervals;
5068 unsigned MaxUsage = 0;
5071 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5072 for (unsigned int i = 0; i < Index; ++i) {
5073 Instruction *I = IdxToInstr[i];
5074 // Ignore instructions that are never used within the loop.
5075 if (!Ends.count(I)) continue;
5077 // Remove all of the instructions that end at this location.
5078 InstrList &List = TransposeEnds[i];
5079 for (unsigned int j=0, e = List.size(); j < e; ++j)
5080 OpenIntervals.erase(List[j]);
5082 // Count the number of live interals.
5083 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5085 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5086 OpenIntervals.size() << '\n');
5088 // Add the current instruction to the list of open intervals.
5089 OpenIntervals.insert(I);
5092 unsigned Invariant = LoopInvariants.size();
5093 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5094 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5095 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5097 R.LoopInvariantRegs = Invariant;
5098 R.MaxLocalUsers = MaxUsage;
5102 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5106 for (Loop::block_iterator bb = TheLoop->block_begin(),
5107 be = TheLoop->block_end(); bb != be; ++bb) {
5108 unsigned BlockCost = 0;
5109 BasicBlock *BB = *bb;
5111 // For each instruction in the old loop.
5112 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5113 // Skip dbg intrinsics.
5114 if (isa<DbgInfoIntrinsic>(it))
5117 unsigned C = getInstructionCost(it, VF);
5119 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5120 VF << " For instruction: " << *it << '\n');
5123 // We assume that if-converted blocks have a 50% chance of being executed.
5124 // When the code is scalar then some of the blocks are avoided due to CF.
5125 // When the code is vectorized we execute all code paths.
5126 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5135 /// \brief Check whether the address computation for a non-consecutive memory
5136 /// access looks like an unlikely candidate for being merged into the indexing
5139 /// We look for a GEP which has one index that is an induction variable and all
5140 /// other indices are loop invariant. If the stride of this access is also
5141 /// within a small bound we decide that this address computation can likely be
5142 /// merged into the addressing mode.
5143 /// In all other cases, we identify the address computation as complex.
5144 static bool isLikelyComplexAddressComputation(Value *Ptr,
5145 LoopVectorizationLegality *Legal,
5146 ScalarEvolution *SE,
5147 const Loop *TheLoop) {
5148 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5152 // We are looking for a gep with all loop invariant indices except for one
5153 // which should be an induction variable.
5154 unsigned NumOperands = Gep->getNumOperands();
5155 for (unsigned i = 1; i < NumOperands; ++i) {
5156 Value *Opd = Gep->getOperand(i);
5157 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5158 !Legal->isInductionVariable(Opd))
5162 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5163 // can likely be merged into the address computation.
5164 unsigned MaxMergeDistance = 64;
5166 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5170 // Check the step is constant.
5171 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5172 // Calculate the pointer stride and check if it is consecutive.
5173 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5177 const APInt &APStepVal = C->getValue()->getValue();
5179 // Huge step value - give up.
5180 if (APStepVal.getBitWidth() > 64)
5183 int64_t StepVal = APStepVal.getSExtValue();
5185 return StepVal > MaxMergeDistance;
5188 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5189 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5195 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5196 // If we know that this instruction will remain uniform, check the cost of
5197 // the scalar version.
5198 if (Legal->isUniformAfterVectorization(I))
5201 Type *RetTy = I->getType();
5202 Type *VectorTy = ToVectorTy(RetTy, VF);
5204 // TODO: We need to estimate the cost of intrinsic calls.
5205 switch (I->getOpcode()) {
5206 case Instruction::GetElementPtr:
5207 // We mark this instruction as zero-cost because the cost of GEPs in
5208 // vectorized code depends on whether the corresponding memory instruction
5209 // is scalarized or not. Therefore, we handle GEPs with the memory
5210 // instruction cost.
5212 case Instruction::Br: {
5213 return TTI.getCFInstrCost(I->getOpcode());
5215 case Instruction::PHI:
5216 //TODO: IF-converted IFs become selects.
5218 case Instruction::Add:
5219 case Instruction::FAdd:
5220 case Instruction::Sub:
5221 case Instruction::FSub:
5222 case Instruction::Mul:
5223 case Instruction::FMul:
5224 case Instruction::UDiv:
5225 case Instruction::SDiv:
5226 case Instruction::FDiv:
5227 case Instruction::URem:
5228 case Instruction::SRem:
5229 case Instruction::FRem:
5230 case Instruction::Shl:
5231 case Instruction::LShr:
5232 case Instruction::AShr:
5233 case Instruction::And:
5234 case Instruction::Or:
5235 case Instruction::Xor: {
5236 // Since we will replace the stride by 1 the multiplication should go away.
5237 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5239 // Certain instructions can be cheaper to vectorize if they have a constant
5240 // second vector operand. One example of this are shifts on x86.
5241 TargetTransformInfo::OperandValueKind Op1VK =
5242 TargetTransformInfo::OK_AnyValue;
5243 TargetTransformInfo::OperandValueKind Op2VK =
5244 TargetTransformInfo::OK_AnyValue;
5246 if (isa<ConstantInt>(I->getOperand(1)))
5247 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5249 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5251 case Instruction::Select: {
5252 SelectInst *SI = cast<SelectInst>(I);
5253 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5254 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5255 Type *CondTy = SI->getCondition()->getType();
5257 CondTy = VectorType::get(CondTy, VF);
5259 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5261 case Instruction::ICmp:
5262 case Instruction::FCmp: {
5263 Type *ValTy = I->getOperand(0)->getType();
5264 VectorTy = ToVectorTy(ValTy, VF);
5265 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5267 case Instruction::Store:
5268 case Instruction::Load: {
5269 StoreInst *SI = dyn_cast<StoreInst>(I);
5270 LoadInst *LI = dyn_cast<LoadInst>(I);
5271 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5273 VectorTy = ToVectorTy(ValTy, VF);
5275 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5276 unsigned AS = SI ? SI->getPointerAddressSpace() :
5277 LI->getPointerAddressSpace();
5278 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5279 // We add the cost of address computation here instead of with the gep
5280 // instruction because only here we know whether the operation is
5283 return TTI.getAddressComputationCost(VectorTy) +
5284 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5286 // Scalarized loads/stores.
5287 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5288 bool Reverse = ConsecutiveStride < 0;
5289 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5290 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5291 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5292 bool IsComplexComputation =
5293 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5295 // The cost of extracting from the value vector and pointer vector.
5296 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5297 for (unsigned i = 0; i < VF; ++i) {
5298 // The cost of extracting the pointer operand.
5299 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5300 // In case of STORE, the cost of ExtractElement from the vector.
5301 // In case of LOAD, the cost of InsertElement into the returned
5303 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5304 Instruction::InsertElement,
5308 // The cost of the scalar loads/stores.
5309 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5310 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5315 // Wide load/stores.
5316 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5317 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5320 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5324 case Instruction::ZExt:
5325 case Instruction::SExt:
5326 case Instruction::FPToUI:
5327 case Instruction::FPToSI:
5328 case Instruction::FPExt:
5329 case Instruction::PtrToInt:
5330 case Instruction::IntToPtr:
5331 case Instruction::SIToFP:
5332 case Instruction::UIToFP:
5333 case Instruction::Trunc:
5334 case Instruction::FPTrunc:
5335 case Instruction::BitCast: {
5336 // We optimize the truncation of induction variable.
5337 // The cost of these is the same as the scalar operation.
5338 if (I->getOpcode() == Instruction::Trunc &&
5339 Legal->isInductionVariable(I->getOperand(0)))
5340 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5341 I->getOperand(0)->getType());
5343 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5344 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5346 case Instruction::Call: {
5347 CallInst *CI = cast<CallInst>(I);
5348 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5349 assert(ID && "Not an intrinsic call!");
5350 Type *RetTy = ToVectorTy(CI->getType(), VF);
5351 SmallVector<Type*, 4> Tys;
5352 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5353 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5354 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5357 // We are scalarizing the instruction. Return the cost of the scalar
5358 // instruction, plus the cost of insert and extract into vector
5359 // elements, times the vector width.
5362 if (!RetTy->isVoidTy() && VF != 1) {
5363 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5365 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5368 // The cost of inserting the results plus extracting each one of the
5370 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5373 // The cost of executing VF copies of the scalar instruction. This opcode
5374 // is unknown. Assume that it is the same as 'mul'.
5375 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5381 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5382 if (Scalar->isVoidTy() || VF == 1)
5384 return VectorType::get(Scalar, VF);
5387 char LoopVectorize::ID = 0;
5388 static const char lv_name[] = "Loop Vectorization";
5389 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5390 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5391 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
5392 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5393 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5394 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5395 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5396 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5399 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5400 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5404 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5405 // Check for a store.
5406 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5407 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5409 // Check for a load.
5410 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5411 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5417 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
5418 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5419 // Holds vector parameters or scalars, in case of uniform vals.
5420 SmallVector<VectorParts, 4> Params;
5422 setDebugLocFromInst(Builder, Instr);
5424 // Find all of the vectorized parameters.
5425 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5426 Value *SrcOp = Instr->getOperand(op);
5428 // If we are accessing the old induction variable, use the new one.
5429 if (SrcOp == OldInduction) {
5430 Params.push_back(getVectorValue(SrcOp));
5434 // Try using previously calculated values.
5435 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5437 // If the src is an instruction that appeared earlier in the basic block
5438 // then it should already be vectorized.
5439 if (SrcInst && OrigLoop->contains(SrcInst)) {
5440 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5441 // The parameter is a vector value from earlier.
5442 Params.push_back(WidenMap.get(SrcInst));
5444 // The parameter is a scalar from outside the loop. Maybe even a constant.
5445 VectorParts Scalars;
5446 Scalars.append(UF, SrcOp);
5447 Params.push_back(Scalars);
5451 assert(Params.size() == Instr->getNumOperands() &&
5452 "Invalid number of operands");
5454 // Does this instruction return a value ?
5455 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5457 Value *UndefVec = IsVoidRetTy ? 0 :
5458 UndefValue::get(Instr->getType());
5459 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5460 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5462 // For each vector unroll 'part':
5463 for (unsigned Part = 0; Part < UF; ++Part) {
5464 // For each scalar that we create:
5466 Instruction *Cloned = Instr->clone();
5468 Cloned->setName(Instr->getName() + ".cloned");
5469 // Replace the operands of the cloned instructions with extracted scalars.
5470 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5471 Value *Op = Params[op][Part];
5472 Cloned->setOperand(op, Op);
5475 // Place the cloned scalar in the new loop.
5476 Builder.Insert(Cloned);
5478 // If the original scalar returns a value we need to place it in a vector
5479 // so that future users will be able to use it.
5481 VecResults[Part] = Cloned;
5485 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5486 return scalarizeInstruction(Instr);
5489 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5493 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5497 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5499 // When unrolling and the VF is 1, we only need to add a simple scalar.
5500 Type *ITy = Val->getType();
5501 assert(!ITy->isVectorTy() && "Val must be a scalar");
5502 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5503 return Builder.CreateAdd(Val, C, "induction");