1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/EquivalenceClasses.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/StringExtras.h"
58 #include "llvm/Analysis/AliasAnalysis.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DebugInfo.h"
71 #include "llvm/IR/DerivedTypes.h"
72 #include "llvm/IR/Dominators.h"
73 #include "llvm/IR/Function.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/Instructions.h"
76 #include "llvm/IR/IntrinsicInst.h"
77 #include "llvm/IR/LLVMContext.h"
78 #include "llvm/IR/Module.h"
79 #include "llvm/IR/PatternMatch.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/Value.h"
82 #include "llvm/IR/ValueHandle.h"
83 #include "llvm/IR/Verifier.h"
84 #include "llvm/Pass.h"
85 #include "llvm/Support/BranchProbability.h"
86 #include "llvm/Support/CommandLine.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/Format.h"
89 #include "llvm/Support/raw_ostream.h"
90 #include "llvm/Target/TargetLibraryInfo.h"
91 #include "llvm/Transforms/Scalar.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/VectorUtils.h"
99 using namespace llvm::PatternMatch;
101 static cl::opt<unsigned>
102 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
103 cl::desc("Sets the SIMD width. Zero is autoselect."));
105 static cl::opt<unsigned>
106 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
107 cl::desc("Sets the vectorization unroll count. "
108 "Zero is autoselect."));
111 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
112 cl::desc("Enable if-conversion during vectorization."));
114 /// We don't vectorize loops with a known constant trip count below this number.
115 static cl::opt<unsigned>
116 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
118 cl::desc("Don't vectorize loops with a constant "
119 "trip count that is smaller than this "
122 /// This enables versioning on the strides of symbolically striding memory
123 /// accesses in code like the following.
124 /// for (i = 0; i < N; ++i)
125 /// A[i * Stride1] += B[i * Stride2] ...
127 /// Will be roughly translated to
128 /// if (Stride1 == 1 && Stride2 == 1) {
129 /// for (i = 0; i < N; i+=4)
133 static cl::opt<bool> EnableMemAccessVersioning(
134 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
135 cl::desc("Enable symblic stride memory access versioning"));
137 /// We don't unroll loops with a known constant trip count below this number.
138 static const unsigned TinyTripCountUnrollThreshold = 128;
140 /// When performing memory disambiguation checks at runtime do not make more
141 /// than this number of comparisons.
142 static const unsigned RuntimeMemoryCheckThreshold = 8;
144 /// Maximum simd width.
145 static const unsigned MaxVectorWidth = 64;
147 static cl::opt<unsigned> ForceTargetNumScalarRegs(
148 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
149 cl::desc("A flag that overrides the target's number of scalar registers."));
151 static cl::opt<unsigned> ForceTargetNumVectorRegs(
152 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
153 cl::desc("A flag that overrides the target's number of vector registers."));
155 /// Maximum vectorization unroll count.
156 static const unsigned MaxUnrollFactor = 16;
158 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
159 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's max unroll factor for scalar "
163 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
164 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's max unroll factor for "
166 "vectorized loops."));
168 static cl::opt<unsigned> ForceTargetInstructionCost(
169 "force-target-instruction-cost", cl::init(0), cl::Hidden,
170 cl::desc("A flag that overrides the target's expected cost for "
171 "an instruction to a single constant value. Mostly "
172 "useful for getting consistent testing."));
174 static cl::opt<unsigned> SmallLoopCost(
175 "small-loop-cost", cl::init(20), cl::Hidden,
176 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
178 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
179 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
180 cl::desc("Enable the use of the block frequency analysis to access PGO "
181 "heuristics minimizing code growth in cold regions and being more "
182 "aggressive in hot regions."));
184 // Runtime unroll loops for load/store throughput.
185 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
186 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
187 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
189 /// The number of stores in a loop that are allowed to need predication.
190 static cl::opt<unsigned> NumberOfStoresToPredicate(
191 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
192 cl::desc("Max number of stores to be predicated behind an if."));
194 static cl::opt<bool> EnableIndVarRegisterHeur(
195 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
196 cl::desc("Count the induction variable only once when unrolling"));
198 static cl::opt<bool> EnableCondStoresVectorization(
199 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
200 cl::desc("Enable if predication of stores during vectorization."));
204 // Forward declarations.
205 class LoopVectorizationLegality;
206 class LoopVectorizationCostModel;
208 /// InnerLoopVectorizer vectorizes loops which contain only one basic
209 /// block to a specified vectorization factor (VF).
210 /// This class performs the widening of scalars into vectors, or multiple
211 /// scalars. This class also implements the following features:
212 /// * It inserts an epilogue loop for handling loops that don't have iteration
213 /// counts that are known to be a multiple of the vectorization factor.
214 /// * It handles the code generation for reduction variables.
215 /// * Scalarization (implementation using scalars) of un-vectorizable
217 /// InnerLoopVectorizer does not perform any vectorization-legality
218 /// checks, and relies on the caller to check for the different legality
219 /// aspects. The InnerLoopVectorizer relies on the
220 /// LoopVectorizationLegality class to provide information about the induction
221 /// and reduction variables that were found to a given vectorization factor.
222 class InnerLoopVectorizer {
224 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
225 DominatorTree *DT, const DataLayout *DL,
226 const TargetLibraryInfo *TLI, unsigned VecWidth,
227 unsigned UnrollFactor)
228 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
229 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
230 OldInduction(0), WidenMap(UnrollFactor), Legal(0) {}
232 // Perform the actual loop widening (vectorization).
233 void vectorize(LoopVectorizationLegality *L) {
235 // Create a new empty loop. Unlink the old loop and connect the new one.
237 // Widen each instruction in the old loop to a new one in the new loop.
238 // Use the Legality module to find the induction and reduction variables.
240 // Register the new loop and update the analysis passes.
244 virtual ~InnerLoopVectorizer() {}
247 /// A small list of PHINodes.
248 typedef SmallVector<PHINode*, 4> PhiVector;
249 /// When we unroll loops we have multiple vector values for each scalar.
250 /// This data structure holds the unrolled and vectorized values that
251 /// originated from one scalar instruction.
252 typedef SmallVector<Value*, 2> VectorParts;
254 // When we if-convert we need create edge masks. We have to cache values so
255 // that we don't end up with exponential recursion/IR.
256 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
257 VectorParts> EdgeMaskCache;
259 /// \brief Add code that checks at runtime if the accessed arrays overlap.
261 /// Returns a pair of instructions where the first element is the first
262 /// instruction generated in possibly a sequence of instructions and the
263 /// second value is the final comparator value or NULL if no check is needed.
264 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
266 /// \brief Add checks for strides that where assumed to be 1.
268 /// Returns the last check instruction and the first check instruction in the
269 /// pair as (first, last).
270 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
272 /// Create an empty loop, based on the loop ranges of the old loop.
273 void createEmptyLoop();
274 /// Copy and widen the instructions from the old loop.
275 virtual void vectorizeLoop();
277 /// \brief The Loop exit block may have single value PHI nodes where the
278 /// incoming value is 'Undef'. While vectorizing we only handled real values
279 /// that were defined inside the loop. Here we fix the 'undef case'.
283 /// A helper function that computes the predicate of the block BB, assuming
284 /// that the header block of the loop is set to True. It returns the *entry*
285 /// mask for the block BB.
286 VectorParts createBlockInMask(BasicBlock *BB);
287 /// A helper function that computes the predicate of the edge between SRC
289 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
291 /// A helper function to vectorize a single BB within the innermost loop.
292 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
294 /// Vectorize a single PHINode in a block. This method handles the induction
295 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
296 /// arbitrary length vectors.
297 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
298 unsigned UF, unsigned VF, PhiVector *PV);
300 /// Insert the new loop to the loop hierarchy and pass manager
301 /// and update the analysis passes.
302 void updateAnalysis();
304 /// This instruction is un-vectorizable. Implement it as a sequence
305 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
306 /// scalarized instruction behind an if block predicated on the control
307 /// dependence of the instruction.
308 virtual void scalarizeInstruction(Instruction *Instr,
309 bool IfPredicateStore=false);
311 /// Vectorize Load and Store instructions,
312 virtual void vectorizeMemoryInstruction(Instruction *Instr);
314 /// Create a broadcast instruction. This method generates a broadcast
315 /// instruction (shuffle) for loop invariant values and for the induction
316 /// value. If this is the induction variable then we extend it to N, N+1, ...
317 /// this is needed because each iteration in the loop corresponds to a SIMD
319 virtual Value *getBroadcastInstrs(Value *V);
321 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
322 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
323 /// The sequence starts at StartIndex.
324 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
326 /// When we go over instructions in the basic block we rely on previous
327 /// values within the current basic block or on loop invariant values.
328 /// When we widen (vectorize) values we place them in the map. If the values
329 /// are not within the map, they have to be loop invariant, so we simply
330 /// broadcast them into a vector.
331 VectorParts &getVectorValue(Value *V);
333 /// Generate a shuffle sequence that will reverse the vector Vec.
334 virtual Value *reverseVector(Value *Vec);
336 /// This is a helper class that holds the vectorizer state. It maps scalar
337 /// instructions to vector instructions. When the code is 'unrolled' then
338 /// then a single scalar value is mapped to multiple vector parts. The parts
339 /// are stored in the VectorPart type.
341 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
343 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
345 /// \return True if 'Key' is saved in the Value Map.
346 bool has(Value *Key) const { return MapStorage.count(Key); }
348 /// Initializes a new entry in the map. Sets all of the vector parts to the
349 /// save value in 'Val'.
350 /// \return A reference to a vector with splat values.
351 VectorParts &splat(Value *Key, Value *Val) {
352 VectorParts &Entry = MapStorage[Key];
353 Entry.assign(UF, Val);
357 ///\return A reference to the value that is stored at 'Key'.
358 VectorParts &get(Value *Key) {
359 VectorParts &Entry = MapStorage[Key];
362 assert(Entry.size() == UF);
367 /// The unroll factor. Each entry in the map stores this number of vector
371 /// Map storage. We use std::map and not DenseMap because insertions to a
372 /// dense map invalidates its iterators.
373 std::map<Value *, VectorParts> MapStorage;
376 /// The original loop.
378 /// Scev analysis to use.
385 const DataLayout *DL;
386 /// Target Library Info.
387 const TargetLibraryInfo *TLI;
389 /// The vectorization SIMD factor to use. Each vector will have this many
394 /// The vectorization unroll factor to use. Each scalar is vectorized to this
395 /// many different vector instructions.
398 /// The builder that we use
401 // --- Vectorization state ---
403 /// The vector-loop preheader.
404 BasicBlock *LoopVectorPreHeader;
405 /// The scalar-loop preheader.
406 BasicBlock *LoopScalarPreHeader;
407 /// Middle Block between the vector and the scalar.
408 BasicBlock *LoopMiddleBlock;
409 ///The ExitBlock of the scalar loop.
410 BasicBlock *LoopExitBlock;
411 ///The vector loop body.
412 SmallVector<BasicBlock *, 4> LoopVectorBody;
413 ///The scalar loop body.
414 BasicBlock *LoopScalarBody;
415 /// A list of all bypass blocks. The first block is the entry of the loop.
416 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
418 /// The new Induction variable which was added to the new block.
420 /// The induction variable of the old basic block.
421 PHINode *OldInduction;
422 /// Holds the extended (to the widest induction type) start index.
424 /// Maps scalars to widened vectors.
426 EdgeMaskCache MaskCache;
428 LoopVectorizationLegality *Legal;
431 class InnerLoopUnroller : public InnerLoopVectorizer {
433 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
434 DominatorTree *DT, const DataLayout *DL,
435 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
436 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
439 void scalarizeInstruction(Instruction *Instr,
440 bool IfPredicateStore = false) override;
441 void vectorizeMemoryInstruction(Instruction *Instr) override;
442 Value *getBroadcastInstrs(Value *V) override;
443 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
444 Value *reverseVector(Value *Vec) override;
447 /// \brief Look for a meaningful debug location on the instruction or it's
449 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
454 if (I->getDebugLoc() != Empty)
457 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
458 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
459 if (OpInst->getDebugLoc() != Empty)
466 /// \brief Set the debug location in the builder using the debug location in the
468 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
469 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
470 B.SetCurrentDebugLocation(Inst->getDebugLoc());
472 B.SetCurrentDebugLocation(DebugLoc());
476 /// \return string containing a file name and a line # for the given
478 static format_object3<const char *, const char *, unsigned>
479 getDebugLocString(const Instruction *I) {
481 return format<const char *, const char *, unsigned>("", "", "", 0U);
482 MDNode *N = I->getMetadata("dbg");
484 const StringRef ModuleName =
485 I->getParent()->getParent()->getParent()->getModuleIdentifier();
486 return format<const char *, const char *, unsigned>("%s", ModuleName.data(),
489 const DILocation Loc(N);
490 const unsigned LineNo = Loc.getLineNumber();
491 const char *DirName = Loc.getDirectory().data();
492 const char *FileName = Loc.getFilename().data();
493 return format("%s/%s:%u", DirName, FileName, LineNo);
497 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
498 /// to what vectorization factor.
499 /// This class does not look at the profitability of vectorization, only the
500 /// legality. This class has two main kinds of checks:
501 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
502 /// will change the order of memory accesses in a way that will change the
503 /// correctness of the program.
504 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
505 /// checks for a number of different conditions, such as the availability of a
506 /// single induction variable, that all types are supported and vectorize-able,
507 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
508 /// This class is also used by InnerLoopVectorizer for identifying
509 /// induction variable and the different reduction variables.
510 class LoopVectorizationLegality {
514 unsigned NumPredStores;
516 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
517 DominatorTree *DT, TargetLibraryInfo *TLI)
518 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
519 DT(DT), TLI(TLI), Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
520 MaxSafeDepDistBytes(-1U) {}
522 /// This enum represents the kinds of reductions that we support.
524 RK_NoReduction, ///< Not a reduction.
525 RK_IntegerAdd, ///< Sum of integers.
526 RK_IntegerMult, ///< Product of integers.
527 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
528 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
529 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
530 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
531 RK_FloatAdd, ///< Sum of floats.
532 RK_FloatMult, ///< Product of floats.
533 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
536 /// This enum represents the kinds of inductions that we support.
538 IK_NoInduction, ///< Not an induction variable.
539 IK_IntInduction, ///< Integer induction variable. Step = 1.
540 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
541 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
542 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
545 // This enum represents the kind of minmax reduction.
546 enum MinMaxReductionKind {
556 /// This struct holds information about reduction variables.
557 struct ReductionDescriptor {
558 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
559 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
561 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
562 MinMaxReductionKind MK)
563 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
565 // The starting value of the reduction.
566 // It does not have to be zero!
567 TrackingVH<Value> StartValue;
568 // The instruction who's value is used outside the loop.
569 Instruction *LoopExitInstr;
570 // The kind of the reduction.
572 // If this a min/max reduction the kind of reduction.
573 MinMaxReductionKind MinMaxKind;
576 /// This POD struct holds information about a potential reduction operation.
577 struct ReductionInstDesc {
578 ReductionInstDesc(bool IsRedux, Instruction *I) :
579 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
581 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
582 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
584 // Is this instruction a reduction candidate.
586 // The last instruction in a min/max pattern (select of the select(icmp())
587 // pattern), or the current reduction instruction otherwise.
588 Instruction *PatternLastInst;
589 // If this is a min/max pattern the comparison predicate.
590 MinMaxReductionKind MinMaxKind;
593 /// This struct holds information about the memory runtime legality
594 /// check that a group of pointers do not overlap.
595 struct RuntimePointerCheck {
596 RuntimePointerCheck() : Need(false) {}
598 /// Reset the state of the pointer runtime information.
605 DependencySetId.clear();
608 /// Insert a pointer and calculate the start and end SCEVs.
609 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
610 unsigned DepSetId, ValueToValueMap &Strides);
612 /// This flag indicates if we need to add the runtime check.
614 /// Holds the pointers that we need to check.
615 SmallVector<TrackingVH<Value>, 2> Pointers;
616 /// Holds the pointer value at the beginning of the loop.
617 SmallVector<const SCEV*, 2> Starts;
618 /// Holds the pointer value at the end of the loop.
619 SmallVector<const SCEV*, 2> Ends;
620 /// Holds the information if this pointer is used for writing to memory.
621 SmallVector<bool, 2> IsWritePtr;
622 /// Holds the id of the set of pointers that could be dependent because of a
623 /// shared underlying object.
624 SmallVector<unsigned, 2> DependencySetId;
627 /// A struct for saving information about induction variables.
628 struct InductionInfo {
629 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
630 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
632 TrackingVH<Value> StartValue;
637 /// ReductionList contains the reduction descriptors for all
638 /// of the reductions that were found in the loop.
639 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
641 /// InductionList saves induction variables and maps them to the
642 /// induction descriptor.
643 typedef MapVector<PHINode*, InductionInfo> InductionList;
645 /// Returns true if it is legal to vectorize this loop.
646 /// This does not mean that it is profitable to vectorize this
647 /// loop, only that it is legal to do so.
650 /// Returns the Induction variable.
651 PHINode *getInduction() { return Induction; }
653 /// Returns the reduction variables found in the loop.
654 ReductionList *getReductionVars() { return &Reductions; }
656 /// Returns the induction variables found in the loop.
657 InductionList *getInductionVars() { return &Inductions; }
659 /// Returns the widest induction type.
660 Type *getWidestInductionType() { return WidestIndTy; }
662 /// Returns True if V is an induction variable in this loop.
663 bool isInductionVariable(const Value *V);
665 /// Return true if the block BB needs to be predicated in order for the loop
666 /// to be vectorized.
667 bool blockNeedsPredication(BasicBlock *BB);
669 /// Check if this pointer is consecutive when vectorizing. This happens
670 /// when the last index of the GEP is the induction variable, or that the
671 /// pointer itself is an induction variable.
672 /// This check allows us to vectorize A[idx] into a wide load/store.
674 /// 0 - Stride is unknown or non-consecutive.
675 /// 1 - Address is consecutive.
676 /// -1 - Address is consecutive, and decreasing.
677 int isConsecutivePtr(Value *Ptr);
679 /// Returns true if the value V is uniform within the loop.
680 bool isUniform(Value *V);
682 /// Returns true if this instruction will remain scalar after vectorization.
683 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
685 /// Returns the information that we collected about runtime memory check.
686 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
688 /// This function returns the identity element (or neutral element) for
690 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
692 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
694 bool hasStride(Value *V) { return StrideSet.count(V); }
695 bool mustCheckStrides() { return !StrideSet.empty(); }
696 SmallPtrSet<Value *, 8>::iterator strides_begin() {
697 return StrideSet.begin();
699 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
702 /// Check if a single basic block loop is vectorizable.
703 /// At this point we know that this is a loop with a constant trip count
704 /// and we only need to check individual instructions.
705 bool canVectorizeInstrs();
707 /// When we vectorize loops we may change the order in which
708 /// we read and write from memory. This method checks if it is
709 /// legal to vectorize the code, considering only memory constrains.
710 /// Returns true if the loop is vectorizable
711 bool canVectorizeMemory();
713 /// Return true if we can vectorize this loop using the IF-conversion
715 bool canVectorizeWithIfConvert();
717 /// Collect the variables that need to stay uniform after vectorization.
718 void collectLoopUniforms();
720 /// Return true if all of the instructions in the block can be speculatively
721 /// executed. \p SafePtrs is a list of addresses that are known to be legal
722 /// and we know that we can read from them without segfault.
723 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
725 /// Returns True, if 'Phi' is the kind of reduction variable for type
726 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
727 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
728 /// Returns a struct describing if the instruction 'I' can be a reduction
729 /// variable of type 'Kind'. If the reduction is a min/max pattern of
730 /// select(icmp()) this function advances the instruction pointer 'I' from the
731 /// compare instruction to the select instruction and stores this pointer in
732 /// 'PatternLastInst' member of the returned struct.
733 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
734 ReductionInstDesc &Desc);
735 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
736 /// pattern corresponding to a min(X, Y) or max(X, Y).
737 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
738 ReductionInstDesc &Prev);
739 /// Returns the induction kind of Phi. This function may return NoInduction
740 /// if the PHI is not an induction variable.
741 InductionKind isInductionVariable(PHINode *Phi);
743 /// \brief Collect memory access with loop invariant strides.
745 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
747 void collectStridedAcccess(Value *LoadOrStoreInst);
749 /// The loop that we evaluate.
753 /// DataLayout analysis.
754 const DataLayout *DL;
757 /// Target Library Info.
758 TargetLibraryInfo *TLI;
760 // --- vectorization state --- //
762 /// Holds the integer induction variable. This is the counter of the
765 /// Holds the reduction variables.
766 ReductionList Reductions;
767 /// Holds all of the induction variables that we found in the loop.
768 /// Notice that inductions don't need to start at zero and that induction
769 /// variables can be pointers.
770 InductionList Inductions;
771 /// Holds the widest induction type encountered.
774 /// Allowed outside users. This holds the reduction
775 /// vars which can be accessed from outside the loop.
776 SmallPtrSet<Value*, 4> AllowedExit;
777 /// This set holds the variables which are known to be uniform after
779 SmallPtrSet<Instruction*, 4> Uniforms;
780 /// We need to check that all of the pointers in this list are disjoint
782 RuntimePointerCheck PtrRtCheck;
783 /// Can we assume the absence of NaNs.
784 bool HasFunNoNaNAttr;
786 unsigned MaxSafeDepDistBytes;
788 ValueToValueMap Strides;
789 SmallPtrSet<Value *, 8> StrideSet;
792 /// LoopVectorizationCostModel - estimates the expected speedups due to
794 /// In many cases vectorization is not profitable. This can happen because of
795 /// a number of reasons. In this class we mainly attempt to predict the
796 /// expected speedup/slowdowns due to the supported instruction set. We use the
797 /// TargetTransformInfo to query the different backends for the cost of
798 /// different operations.
799 class LoopVectorizationCostModel {
801 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
802 LoopVectorizationLegality *Legal,
803 const TargetTransformInfo &TTI,
804 const DataLayout *DL, const TargetLibraryInfo *TLI)
805 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
807 /// Information about vectorization costs
808 struct VectorizationFactor {
809 unsigned Width; // Vector width with best cost
810 unsigned Cost; // Cost of the loop with that width
812 /// \return The most profitable vectorization factor and the cost of that VF.
813 /// This method checks every power of two up to VF. If UserVF is not ZERO
814 /// then this vectorization factor will be selected if vectorization is
816 VectorizationFactor selectVectorizationFactor(bool OptForSize,
819 /// \return The size (in bits) of the widest type in the code that
820 /// needs to be vectorized. We ignore values that remain scalar such as
821 /// 64 bit loop indices.
822 unsigned getWidestType();
824 /// \return The most profitable unroll factor.
825 /// If UserUF is non-zero then this method finds the best unroll-factor
826 /// based on register pressure and other parameters.
827 /// VF and LoopCost are the selected vectorization factor and the cost of the
829 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
832 /// \brief A struct that represents some properties of the register usage
834 struct RegisterUsage {
835 /// Holds the number of loop invariant values that are used in the loop.
836 unsigned LoopInvariantRegs;
837 /// Holds the maximum number of concurrent live intervals in the loop.
838 unsigned MaxLocalUsers;
839 /// Holds the number of instructions in the loop.
840 unsigned NumInstructions;
843 /// \return information about the register usage of the loop.
844 RegisterUsage calculateRegisterUsage();
847 /// Returns the expected execution cost. The unit of the cost does
848 /// not matter because we use the 'cost' units to compare different
849 /// vector widths. The cost that is returned is *not* normalized by
850 /// the factor width.
851 unsigned expectedCost(unsigned VF);
853 /// Returns the execution time cost of an instruction for a given vector
854 /// width. Vector width of one means scalar.
855 unsigned getInstructionCost(Instruction *I, unsigned VF);
857 /// A helper function for converting Scalar types to vector types.
858 /// If the incoming type is void, we return void. If the VF is 1, we return
860 static Type* ToVectorTy(Type *Scalar, unsigned VF);
862 /// Returns whether the instruction is a load or store and will be a emitted
863 /// as a vector operation.
864 bool isConsecutiveLoadOrStore(Instruction *I);
866 /// The loop that we evaluate.
870 /// Loop Info analysis.
872 /// Vectorization legality.
873 LoopVectorizationLegality *Legal;
874 /// Vector target information.
875 const TargetTransformInfo &TTI;
876 /// Target data layout information.
877 const DataLayout *DL;
878 /// Target Library Info.
879 const TargetLibraryInfo *TLI;
882 /// Utility class for getting and setting loop vectorizer hints in the form
883 /// of loop metadata.
884 struct LoopVectorizeHints {
885 /// Vectorization width.
887 /// Vectorization unroll factor.
889 /// Vectorization forced (-1 not selected, 0 force disabled, 1 force enabled)
892 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
893 : Width(VectorizationFactor)
894 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
896 , LoopID(L->getLoopID()) {
898 // The command line options override any loop metadata except for when
899 // width == 1 which is used to indicate the loop is already vectorized.
900 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
901 Width = VectorizationFactor;
902 if (VectorizationUnroll.getNumOccurrences() > 0)
903 Unroll = VectorizationUnroll;
905 DEBUG(if (DisableUnrolling && Unroll == 1)
906 dbgs() << "LV: Unrolling disabled by the pass manager\n");
909 /// Return the loop vectorizer metadata prefix.
910 static StringRef Prefix() { return "llvm.vectorizer."; }
912 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
913 SmallVector<Value*, 2> Vals;
914 Vals.push_back(MDString::get(Context, Name));
915 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
916 return MDNode::get(Context, Vals);
919 /// Mark the loop L as already vectorized by setting the width to 1.
920 void setAlreadyVectorized(Loop *L) {
921 LLVMContext &Context = L->getHeader()->getContext();
925 // Create a new loop id with one more operand for the already_vectorized
926 // hint. If the loop already has a loop id then copy the existing operands.
927 SmallVector<Value*, 4> Vals(1);
929 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
930 Vals.push_back(LoopID->getOperand(i));
932 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
933 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
935 MDNode *NewLoopID = MDNode::get(Context, Vals);
936 // Set operand 0 to refer to the loop id itself.
937 NewLoopID->replaceOperandWith(0, NewLoopID);
939 L->setLoopID(NewLoopID);
941 LoopID->replaceAllUsesWith(NewLoopID);
949 /// Find hints specified in the loop metadata.
950 void getHints(const Loop *L) {
954 // First operand should refer to the loop id itself.
955 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
956 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
958 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
959 const MDString *S = 0;
960 SmallVector<Value*, 4> Args;
962 // The expected hint is either a MDString or a MDNode with the first
963 // operand a MDString.
964 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
965 if (!MD || MD->getNumOperands() == 0)
967 S = dyn_cast<MDString>(MD->getOperand(0));
968 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
969 Args.push_back(MD->getOperand(i));
971 S = dyn_cast<MDString>(LoopID->getOperand(i));
972 assert(Args.size() == 0 && "too many arguments for MDString");
978 // Check if the hint starts with the vectorizer prefix.
979 StringRef Hint = S->getString();
980 if (!Hint.startswith(Prefix()))
982 // Remove the prefix.
983 Hint = Hint.substr(Prefix().size(), StringRef::npos);
985 if (Args.size() == 1)
986 getHint(Hint, Args[0]);
990 // Check string hint with one operand.
991 void getHint(StringRef Hint, Value *Arg) {
992 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
994 unsigned Val = C->getZExtValue();
996 if (Hint == "width") {
997 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1000 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1001 } else if (Hint == "unroll") {
1002 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1005 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1006 } else if (Hint == "enable") {
1007 if (C->getBitWidth() == 1)
1010 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1012 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1017 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1019 return V.push_back(&L);
1021 for (Loop *InnerL : L)
1022 addInnerLoop(*InnerL, V);
1025 /// The LoopVectorize Pass.
1026 struct LoopVectorize : public FunctionPass {
1027 /// Pass identification, replacement for typeid
1030 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1032 DisableUnrolling(NoUnrolling),
1033 AlwaysVectorize(AlwaysVectorize) {
1034 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1037 ScalarEvolution *SE;
1038 const DataLayout *DL;
1040 TargetTransformInfo *TTI;
1042 BlockFrequencyInfo *BFI;
1043 TargetLibraryInfo *TLI;
1044 bool DisableUnrolling;
1045 bool AlwaysVectorize;
1047 BlockFrequency ColdEntryFreq;
1049 bool runOnFunction(Function &F) override {
1050 SE = &getAnalysis<ScalarEvolution>();
1051 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1052 DL = DLP ? &DLP->getDataLayout() : 0;
1053 LI = &getAnalysis<LoopInfo>();
1054 TTI = &getAnalysis<TargetTransformInfo>();
1055 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1056 BFI = &getAnalysis<BlockFrequencyInfo>();
1057 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1059 // Compute some weights outside of the loop over the loops. Compute this
1060 // using a BranchProbability to re-use its scaling math.
1061 const BranchProbability ColdProb(1, 5); // 20%
1062 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1064 // If the target claims to have no vector registers don't attempt
1066 if (!TTI->getNumberOfRegisters(true))
1070 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1071 << ": Missing data layout\n");
1075 // Build up a worklist of inner-loops to vectorize. This is necessary as
1076 // the act of vectorizing or partially unrolling a loop creates new loops
1077 // and can invalidate iterators across the loops.
1078 SmallVector<Loop *, 8> Worklist;
1081 addInnerLoop(*L, Worklist);
1083 // Now walk the identified inner loops.
1084 bool Changed = false;
1085 while (!Worklist.empty())
1086 Changed |= processLoop(Worklist.pop_back_val());
1088 // Process each loop nest in the function.
1092 bool processLoop(Loop *L) {
1093 assert(L->empty() && "Only process inner loops.");
1094 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1095 << L->getHeader()->getParent()->getName() << "\" from "
1096 << getDebugLocString(L->getHeader()->getFirstNonPHIOrDbg())
1099 LoopVectorizeHints Hints(L, DisableUnrolling);
1101 DEBUG(dbgs() << "LV: Loop hints:"
1102 << " force=" << (Hints.Force == 0
1104 : (Hints.Force == 1 ? "enabled" : "?"))
1105 << " width=" << Hints.Width << " unroll=" << Hints.Unroll
1108 if (Hints.Force == 0) {
1109 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1113 if (!AlwaysVectorize && Hints.Force != 1) {
1114 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1118 if (Hints.Width == 1 && Hints.Unroll == 1) {
1119 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1123 // Check if it is legal to vectorize the loop.
1124 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1125 if (!LVL.canVectorize()) {
1126 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1130 // Use the cost model.
1131 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1133 // Check the function attributes to find out if this function should be
1134 // optimized for size.
1135 Function *F = L->getHeader()->getParent();
1137 Hints.Force != 1 && F->hasFnAttribute(Attribute::OptimizeForSize);
1139 // Compute the weighted frequency of this loop being executed and see if it
1140 // is less than 20% of the function entry baseline frequency. Note that we
1141 // always have a canonical loop here because we think we *can* vectoriez.
1142 // FIXME: This is hidden behind a flag due to pervasive problems with
1143 // exactly what block frequency models.
1144 if (LoopVectorizeWithBlockFrequency) {
1145 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1146 if (Hints.Force != 1 && LoopEntryFreq < ColdEntryFreq)
1150 // Check the function attributes to see if implicit floats are allowed.a
1151 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1152 // an integer loop and the vector instructions selected are purely integer
1153 // vector instructions?
1154 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1155 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1156 "attribute is used.\n");
1160 // Select the optimal vectorization factor.
1161 const LoopVectorizationCostModel::VectorizationFactor VF =
1162 CM.selectVectorizationFactor(OptForSize, Hints.Width);
1163 // Select the unroll factor.
1164 const unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1167 DEBUG(dbgs() << "LV: Found a vectorizable loop ("
1168 << VF.Width << ") in "
1169 << getDebugLocString(L->getHeader()->getFirstNonPHIOrDbg())
1171 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1173 if (VF.Width == 1) {
1174 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1177 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1178 // We decided not to vectorize, but we may want to unroll.
1179 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1180 Unroller.vectorize(&LVL);
1182 // If we decided that it is *legal* to vectorize the loop then do it.
1183 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1187 // Mark the loop as already vectorized to avoid vectorizing again.
1188 Hints.setAlreadyVectorized(L);
1190 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1194 void getAnalysisUsage(AnalysisUsage &AU) const override {
1195 AU.addRequiredID(LoopSimplifyID);
1196 AU.addRequiredID(LCSSAID);
1197 AU.addRequired<BlockFrequencyInfo>();
1198 AU.addRequired<DominatorTreeWrapperPass>();
1199 AU.addRequired<LoopInfo>();
1200 AU.addRequired<ScalarEvolution>();
1201 AU.addRequired<TargetTransformInfo>();
1202 AU.addPreserved<LoopInfo>();
1203 AU.addPreserved<DominatorTreeWrapperPass>();
1208 } // end anonymous namespace
1210 //===----------------------------------------------------------------------===//
1211 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1212 // LoopVectorizationCostModel.
1213 //===----------------------------------------------------------------------===//
1215 static Value *stripIntegerCast(Value *V) {
1216 if (CastInst *CI = dyn_cast<CastInst>(V))
1217 if (CI->getOperand(0)->getType()->isIntegerTy())
1218 return CI->getOperand(0);
1222 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1224 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1226 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1227 ValueToValueMap &PtrToStride,
1228 Value *Ptr, Value *OrigPtr = 0) {
1230 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1232 // If there is an entry in the map return the SCEV of the pointer with the
1233 // symbolic stride replaced by one.
1234 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1235 if (SI != PtrToStride.end()) {
1236 Value *StrideVal = SI->second;
1239 StrideVal = stripIntegerCast(StrideVal);
1241 // Replace symbolic stride by one.
1242 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1243 ValueToValueMap RewriteMap;
1244 RewriteMap[StrideVal] = One;
1247 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1248 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1253 // Otherwise, just return the SCEV of the original pointer.
1254 return SE->getSCEV(Ptr);
1257 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1258 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1259 ValueToValueMap &Strides) {
1260 // Get the stride replaced scev.
1261 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1262 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1263 assert(AR && "Invalid addrec expression");
1264 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1265 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1266 Pointers.push_back(Ptr);
1267 Starts.push_back(AR->getStart());
1268 Ends.push_back(ScEnd);
1269 IsWritePtr.push_back(WritePtr);
1270 DependencySetId.push_back(DepSetId);
1273 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1274 // We need to place the broadcast of invariant variables outside the loop.
1275 Instruction *Instr = dyn_cast<Instruction>(V);
1277 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1278 Instr->getParent()) != LoopVectorBody.end());
1279 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1281 // Place the code for broadcasting invariant variables in the new preheader.
1282 IRBuilder<>::InsertPointGuard Guard(Builder);
1284 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1286 // Broadcast the scalar into all locations in the vector.
1287 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1292 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1294 assert(Val->getType()->isVectorTy() && "Must be a vector");
1295 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1296 "Elem must be an integer");
1297 // Create the types.
1298 Type *ITy = Val->getType()->getScalarType();
1299 VectorType *Ty = cast<VectorType>(Val->getType());
1300 int VLen = Ty->getNumElements();
1301 SmallVector<Constant*, 8> Indices;
1303 // Create a vector of consecutive numbers from zero to VF.
1304 for (int i = 0; i < VLen; ++i) {
1305 int64_t Idx = Negate ? (-i) : i;
1306 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1309 // Add the consecutive indices to the vector value.
1310 Constant *Cv = ConstantVector::get(Indices);
1311 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1312 return Builder.CreateAdd(Val, Cv, "induction");
1315 /// \brief Find the operand of the GEP that should be checked for consecutive
1316 /// stores. This ignores trailing indices that have no effect on the final
1318 static unsigned getGEPInductionOperand(const DataLayout *DL,
1319 const GetElementPtrInst *Gep) {
1320 unsigned LastOperand = Gep->getNumOperands() - 1;
1321 unsigned GEPAllocSize = DL->getTypeAllocSize(
1322 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1324 // Walk backwards and try to peel off zeros.
1325 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1326 // Find the type we're currently indexing into.
1327 gep_type_iterator GEPTI = gep_type_begin(Gep);
1328 std::advance(GEPTI, LastOperand - 1);
1330 // If it's a type with the same allocation size as the result of the GEP we
1331 // can peel off the zero index.
1332 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1340 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1341 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1342 // Make sure that the pointer does not point to structs.
1343 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1346 // If this value is a pointer induction variable we know it is consecutive.
1347 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1348 if (Phi && Inductions.count(Phi)) {
1349 InductionInfo II = Inductions[Phi];
1350 if (IK_PtrInduction == II.IK)
1352 else if (IK_ReversePtrInduction == II.IK)
1356 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1360 unsigned NumOperands = Gep->getNumOperands();
1361 Value *GpPtr = Gep->getPointerOperand();
1362 // If this GEP value is a consecutive pointer induction variable and all of
1363 // the indices are constant then we know it is consecutive. We can
1364 Phi = dyn_cast<PHINode>(GpPtr);
1365 if (Phi && Inductions.count(Phi)) {
1367 // Make sure that the pointer does not point to structs.
1368 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1369 if (GepPtrType->getElementType()->isAggregateType())
1372 // Make sure that all of the index operands are loop invariant.
1373 for (unsigned i = 1; i < NumOperands; ++i)
1374 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1377 InductionInfo II = Inductions[Phi];
1378 if (IK_PtrInduction == II.IK)
1380 else if (IK_ReversePtrInduction == II.IK)
1384 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1386 // Check that all of the gep indices are uniform except for our induction
1388 for (unsigned i = 0; i != NumOperands; ++i)
1389 if (i != InductionOperand &&
1390 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1393 // We can emit wide load/stores only if the last non-zero index is the
1394 // induction variable.
1395 const SCEV *Last = 0;
1396 if (!Strides.count(Gep))
1397 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1399 // Because of the multiplication by a stride we can have a s/zext cast.
1400 // We are going to replace this stride by 1 so the cast is safe to ignore.
1402 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1403 // %0 = trunc i64 %indvars.iv to i32
1404 // %mul = mul i32 %0, %Stride1
1405 // %idxprom = zext i32 %mul to i64 << Safe cast.
1406 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1408 Last = replaceSymbolicStrideSCEV(SE, Strides,
1409 Gep->getOperand(InductionOperand), Gep);
1410 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1412 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1416 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1417 const SCEV *Step = AR->getStepRecurrence(*SE);
1419 // The memory is consecutive because the last index is consecutive
1420 // and all other indices are loop invariant.
1423 if (Step->isAllOnesValue())
1430 bool LoopVectorizationLegality::isUniform(Value *V) {
1431 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1434 InnerLoopVectorizer::VectorParts&
1435 InnerLoopVectorizer::getVectorValue(Value *V) {
1436 assert(V != Induction && "The new induction variable should not be used.");
1437 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1439 // If we have a stride that is replaced by one, do it here.
1440 if (Legal->hasStride(V))
1441 V = ConstantInt::get(V->getType(), 1);
1443 // If we have this scalar in the map, return it.
1444 if (WidenMap.has(V))
1445 return WidenMap.get(V);
1447 // If this scalar is unknown, assume that it is a constant or that it is
1448 // loop invariant. Broadcast V and save the value for future uses.
1449 Value *B = getBroadcastInstrs(V);
1450 return WidenMap.splat(V, B);
1453 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1454 assert(Vec->getType()->isVectorTy() && "Invalid type");
1455 SmallVector<Constant*, 8> ShuffleMask;
1456 for (unsigned i = 0; i < VF; ++i)
1457 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1459 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1460 ConstantVector::get(ShuffleMask),
1464 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1465 // Attempt to issue a wide load.
1466 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1467 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1469 assert((LI || SI) && "Invalid Load/Store instruction");
1471 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1472 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1473 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1474 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1475 // An alignment of 0 means target abi alignment. We need to use the scalar's
1476 // target abi alignment in such a case.
1478 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1479 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1480 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1481 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1483 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1484 return scalarizeInstruction(Instr, true);
1486 if (ScalarAllocatedSize != VectorElementSize)
1487 return scalarizeInstruction(Instr);
1489 // If the pointer is loop invariant or if it is non-consecutive,
1490 // scalarize the load.
1491 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1492 bool Reverse = ConsecutiveStride < 0;
1493 bool UniformLoad = LI && Legal->isUniform(Ptr);
1494 if (!ConsecutiveStride || UniformLoad)
1495 return scalarizeInstruction(Instr);
1497 Constant *Zero = Builder.getInt32(0);
1498 VectorParts &Entry = WidenMap.get(Instr);
1500 // Handle consecutive loads/stores.
1501 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1502 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1503 setDebugLocFromInst(Builder, Gep);
1504 Value *PtrOperand = Gep->getPointerOperand();
1505 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1506 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1508 // Create the new GEP with the new induction variable.
1509 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1510 Gep2->setOperand(0, FirstBasePtr);
1511 Gep2->setName("gep.indvar.base");
1512 Ptr = Builder.Insert(Gep2);
1514 setDebugLocFromInst(Builder, Gep);
1515 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1516 OrigLoop) && "Base ptr must be invariant");
1518 // The last index does not have to be the induction. It can be
1519 // consecutive and be a function of the index. For example A[I+1];
1520 unsigned NumOperands = Gep->getNumOperands();
1521 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1522 // Create the new GEP with the new induction variable.
1523 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1525 for (unsigned i = 0; i < NumOperands; ++i) {
1526 Value *GepOperand = Gep->getOperand(i);
1527 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1529 // Update last index or loop invariant instruction anchored in loop.
1530 if (i == InductionOperand ||
1531 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1532 assert((i == InductionOperand ||
1533 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1534 "Must be last index or loop invariant");
1536 VectorParts &GEPParts = getVectorValue(GepOperand);
1537 Value *Index = GEPParts[0];
1538 Index = Builder.CreateExtractElement(Index, Zero);
1539 Gep2->setOperand(i, Index);
1540 Gep2->setName("gep.indvar.idx");
1543 Ptr = Builder.Insert(Gep2);
1545 // Use the induction element ptr.
1546 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1547 setDebugLocFromInst(Builder, Ptr);
1548 VectorParts &PtrVal = getVectorValue(Ptr);
1549 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1554 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1555 "We do not allow storing to uniform addresses");
1556 setDebugLocFromInst(Builder, SI);
1557 // We don't want to update the value in the map as it might be used in
1558 // another expression. So don't use a reference type for "StoredVal".
1559 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1561 for (unsigned Part = 0; Part < UF; ++Part) {
1562 // Calculate the pointer for the specific unroll-part.
1563 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1566 // If we store to reverse consecutive memory locations then we need
1567 // to reverse the order of elements in the stored value.
1568 StoredVal[Part] = reverseVector(StoredVal[Part]);
1569 // If the address is consecutive but reversed, then the
1570 // wide store needs to start at the last vector element.
1571 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1572 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1575 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1576 DataTy->getPointerTo(AddressSpace));
1577 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1583 assert(LI && "Must have a load instruction");
1584 setDebugLocFromInst(Builder, LI);
1585 for (unsigned Part = 0; Part < UF; ++Part) {
1586 // Calculate the pointer for the specific unroll-part.
1587 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1590 // If the address is consecutive but reversed, then the
1591 // wide store needs to start at the last vector element.
1592 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1593 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1596 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1597 DataTy->getPointerTo(AddressSpace));
1598 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1599 cast<LoadInst>(LI)->setAlignment(Alignment);
1600 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1604 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1605 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1606 // Holds vector parameters or scalars, in case of uniform vals.
1607 SmallVector<VectorParts, 4> Params;
1609 setDebugLocFromInst(Builder, Instr);
1611 // Find all of the vectorized parameters.
1612 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1613 Value *SrcOp = Instr->getOperand(op);
1615 // If we are accessing the old induction variable, use the new one.
1616 if (SrcOp == OldInduction) {
1617 Params.push_back(getVectorValue(SrcOp));
1621 // Try using previously calculated values.
1622 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1624 // If the src is an instruction that appeared earlier in the basic block
1625 // then it should already be vectorized.
1626 if (SrcInst && OrigLoop->contains(SrcInst)) {
1627 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1628 // The parameter is a vector value from earlier.
1629 Params.push_back(WidenMap.get(SrcInst));
1631 // The parameter is a scalar from outside the loop. Maybe even a constant.
1632 VectorParts Scalars;
1633 Scalars.append(UF, SrcOp);
1634 Params.push_back(Scalars);
1638 assert(Params.size() == Instr->getNumOperands() &&
1639 "Invalid number of operands");
1641 // Does this instruction return a value ?
1642 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1644 Value *UndefVec = IsVoidRetTy ? 0 :
1645 UndefValue::get(VectorType::get(Instr->getType(), VF));
1646 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1647 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1649 Instruction *InsertPt = Builder.GetInsertPoint();
1650 BasicBlock *IfBlock = Builder.GetInsertBlock();
1651 BasicBlock *CondBlock = 0;
1655 if (IfPredicateStore) {
1656 assert(Instr->getParent()->getSinglePredecessor() &&
1657 "Only support single predecessor blocks");
1658 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1659 Instr->getParent());
1660 VectorLp = LI->getLoopFor(IfBlock);
1661 assert(VectorLp && "Must have a loop for this block");
1664 // For each vector unroll 'part':
1665 for (unsigned Part = 0; Part < UF; ++Part) {
1666 // For each scalar that we create:
1667 for (unsigned Width = 0; Width < VF; ++Width) {
1671 if (IfPredicateStore) {
1672 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1673 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1674 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1675 LoopVectorBody.push_back(CondBlock);
1676 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1677 // Update Builder with newly created basic block.
1678 Builder.SetInsertPoint(InsertPt);
1681 Instruction *Cloned = Instr->clone();
1683 Cloned->setName(Instr->getName() + ".cloned");
1684 // Replace the operands of the cloned instructions with extracted scalars.
1685 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1686 Value *Op = Params[op][Part];
1687 // Param is a vector. Need to extract the right lane.
1688 if (Op->getType()->isVectorTy())
1689 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1690 Cloned->setOperand(op, Op);
1693 // Place the cloned scalar in the new loop.
1694 Builder.Insert(Cloned);
1696 // If the original scalar returns a value we need to place it in a vector
1697 // so that future users will be able to use it.
1699 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1700 Builder.getInt32(Width));
1702 if (IfPredicateStore) {
1703 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1704 LoopVectorBody.push_back(NewIfBlock);
1705 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1706 Builder.SetInsertPoint(InsertPt);
1707 Instruction *OldBr = IfBlock->getTerminator();
1708 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1709 OldBr->eraseFromParent();
1710 IfBlock = NewIfBlock;
1716 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1720 if (Instruction *I = dyn_cast<Instruction>(V))
1721 return I->getParent() == Loc->getParent() ? I : 0;
1725 std::pair<Instruction *, Instruction *>
1726 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1727 Instruction *tnullptr = 0;
1728 if (!Legal->mustCheckStrides())
1729 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1731 IRBuilder<> ChkBuilder(Loc);
1735 Instruction *FirstInst = 0;
1736 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1737 SE = Legal->strides_end();
1739 Value *Ptr = stripIntegerCast(*SI);
1740 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1742 // Store the first instruction we create.
1743 FirstInst = getFirstInst(FirstInst, C, Loc);
1745 Check = ChkBuilder.CreateOr(Check, C);
1750 // We have to do this trickery because the IRBuilder might fold the check to a
1751 // constant expression in which case there is no Instruction anchored in a
1753 LLVMContext &Ctx = Loc->getContext();
1754 Instruction *TheCheck =
1755 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1756 ChkBuilder.Insert(TheCheck, "stride.not.one");
1757 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1759 return std::make_pair(FirstInst, TheCheck);
1762 std::pair<Instruction *, Instruction *>
1763 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1764 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1765 Legal->getRuntimePointerCheck();
1767 Instruction *tnullptr = 0;
1768 if (!PtrRtCheck->Need)
1769 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1771 unsigned NumPointers = PtrRtCheck->Pointers.size();
1772 SmallVector<TrackingVH<Value> , 2> Starts;
1773 SmallVector<TrackingVH<Value> , 2> Ends;
1775 LLVMContext &Ctx = Loc->getContext();
1776 SCEVExpander Exp(*SE, "induction");
1777 Instruction *FirstInst = 0;
1779 for (unsigned i = 0; i < NumPointers; ++i) {
1780 Value *Ptr = PtrRtCheck->Pointers[i];
1781 const SCEV *Sc = SE->getSCEV(Ptr);
1783 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1784 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1786 Starts.push_back(Ptr);
1787 Ends.push_back(Ptr);
1789 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1790 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1792 // Use this type for pointer arithmetic.
1793 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1795 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1796 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1797 Starts.push_back(Start);
1798 Ends.push_back(End);
1802 IRBuilder<> ChkBuilder(Loc);
1803 // Our instructions might fold to a constant.
1804 Value *MemoryRuntimeCheck = 0;
1805 for (unsigned i = 0; i < NumPointers; ++i) {
1806 for (unsigned j = i+1; j < NumPointers; ++j) {
1807 // No need to check if two readonly pointers intersect.
1808 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1811 // Only need to check pointers between two different dependency sets.
1812 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1815 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1816 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1818 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1819 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1820 "Trying to bounds check pointers with different address spaces");
1822 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1823 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1825 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1826 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1827 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1828 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1830 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1831 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1832 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1833 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1834 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1835 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1836 if (MemoryRuntimeCheck) {
1837 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1839 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1841 MemoryRuntimeCheck = IsConflict;
1845 // We have to do this trickery because the IRBuilder might fold the check to a
1846 // constant expression in which case there is no Instruction anchored in a
1848 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1849 ConstantInt::getTrue(Ctx));
1850 ChkBuilder.Insert(Check, "memcheck.conflict");
1851 FirstInst = getFirstInst(FirstInst, Check, Loc);
1852 return std::make_pair(FirstInst, Check);
1855 void InnerLoopVectorizer::createEmptyLoop() {
1857 In this function we generate a new loop. The new loop will contain
1858 the vectorized instructions while the old loop will continue to run the
1861 [ ] <-- vector loop bypass (may consist of multiple blocks).
1864 | [ ] <-- vector pre header.
1868 | [ ]_| <-- vector loop.
1871 >[ ] <--- middle-block.
1874 | [ ] <--- new preheader.
1878 | [ ]_| <-- old scalar loop to handle remainder.
1881 >[ ] <-- exit block.
1885 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1886 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1887 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1888 assert(ExitBlock && "Must have an exit block");
1890 // Some loops have a single integer induction variable, while other loops
1891 // don't. One example is c++ iterators that often have multiple pointer
1892 // induction variables. In the code below we also support a case where we
1893 // don't have a single induction variable.
1894 OldInduction = Legal->getInduction();
1895 Type *IdxTy = Legal->getWidestInductionType();
1897 // Find the loop boundaries.
1898 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1899 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1901 // The exit count might have the type of i64 while the phi is i32. This can
1902 // happen if we have an induction variable that is sign extended before the
1903 // compare. The only way that we get a backedge taken count is that the
1904 // induction variable was signed and as such will not overflow. In such a case
1905 // truncation is legal.
1906 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1907 IdxTy->getPrimitiveSizeInBits())
1908 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1910 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1911 // Get the total trip count from the count by adding 1.
1912 ExitCount = SE->getAddExpr(ExitCount,
1913 SE->getConstant(ExitCount->getType(), 1));
1915 // Expand the trip count and place the new instructions in the preheader.
1916 // Notice that the pre-header does not change, only the loop body.
1917 SCEVExpander Exp(*SE, "induction");
1919 // Count holds the overall loop count (N).
1920 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1921 BypassBlock->getTerminator());
1923 // The loop index does not have to start at Zero. Find the original start
1924 // value from the induction PHI node. If we don't have an induction variable
1925 // then we know that it starts at zero.
1926 Builder.SetInsertPoint(BypassBlock->getTerminator());
1927 Value *StartIdx = ExtendedIdx = OldInduction ?
1928 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1930 ConstantInt::get(IdxTy, 0);
1932 assert(BypassBlock && "Invalid loop structure");
1933 LoopBypassBlocks.push_back(BypassBlock);
1935 // Split the single block loop into the two loop structure described above.
1936 BasicBlock *VectorPH =
1937 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1938 BasicBlock *VecBody =
1939 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1940 BasicBlock *MiddleBlock =
1941 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1942 BasicBlock *ScalarPH =
1943 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1945 // Create and register the new vector loop.
1946 Loop* Lp = new Loop();
1947 Loop *ParentLoop = OrigLoop->getParentLoop();
1949 // Insert the new loop into the loop nest and register the new basic blocks
1950 // before calling any utilities such as SCEV that require valid LoopInfo.
1952 ParentLoop->addChildLoop(Lp);
1953 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1954 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1955 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1957 LI->addTopLevelLoop(Lp);
1959 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1961 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1963 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1965 // Generate the induction variable.
1966 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1967 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1968 // The loop step is equal to the vectorization factor (num of SIMD elements)
1969 // times the unroll factor (num of SIMD instructions).
1970 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1972 // This is the IR builder that we use to add all of the logic for bypassing
1973 // the new vector loop.
1974 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1975 setDebugLocFromInst(BypassBuilder,
1976 getDebugLocFromInstOrOperands(OldInduction));
1978 // We may need to extend the index in case there is a type mismatch.
1979 // We know that the count starts at zero and does not overflow.
1980 if (Count->getType() != IdxTy) {
1981 // The exit count can be of pointer type. Convert it to the correct
1983 if (ExitCount->getType()->isPointerTy())
1984 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1986 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1989 // Add the start index to the loop count to get the new end index.
1990 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1992 // Now we need to generate the expression for N - (N % VF), which is
1993 // the part that the vectorized body will execute.
1994 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1995 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1996 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1997 "end.idx.rnd.down");
1999 // Now, compare the new count to zero. If it is zero skip the vector loop and
2000 // jump to the scalar loop.
2001 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
2004 BasicBlock *LastBypassBlock = BypassBlock;
2006 // Generate the code to check that the strides we assumed to be one are really
2007 // one. We want the new basic block to start at the first instruction in a
2008 // sequence of instructions that form a check.
2009 Instruction *StrideCheck;
2010 Instruction *FirstCheckInst;
2011 std::tie(FirstCheckInst, StrideCheck) =
2012 addStrideCheck(BypassBlock->getTerminator());
2014 // Create a new block containing the stride check.
2015 BasicBlock *CheckBlock =
2016 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2018 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2019 LoopBypassBlocks.push_back(CheckBlock);
2021 // Replace the branch into the memory check block with a conditional branch
2022 // for the "few elements case".
2023 Instruction *OldTerm = BypassBlock->getTerminator();
2024 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2025 OldTerm->eraseFromParent();
2028 LastBypassBlock = CheckBlock;
2031 // Generate the code that checks in runtime if arrays overlap. We put the
2032 // checks into a separate block to make the more common case of few elements
2034 Instruction *MemRuntimeCheck;
2035 std::tie(FirstCheckInst, MemRuntimeCheck) =
2036 addRuntimeCheck(LastBypassBlock->getTerminator());
2037 if (MemRuntimeCheck) {
2038 // Create a new block containing the memory check.
2039 BasicBlock *CheckBlock =
2040 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2042 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2043 LoopBypassBlocks.push_back(CheckBlock);
2045 // Replace the branch into the memory check block with a conditional branch
2046 // for the "few elements case".
2047 Instruction *OldTerm = LastBypassBlock->getTerminator();
2048 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2049 OldTerm->eraseFromParent();
2051 Cmp = MemRuntimeCheck;
2052 LastBypassBlock = CheckBlock;
2055 LastBypassBlock->getTerminator()->eraseFromParent();
2056 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2059 // We are going to resume the execution of the scalar loop.
2060 // Go over all of the induction variables that we found and fix the
2061 // PHIs that are left in the scalar version of the loop.
2062 // The starting values of PHI nodes depend on the counter of the last
2063 // iteration in the vectorized loop.
2064 // If we come from a bypass edge then we need to start from the original
2067 // This variable saves the new starting index for the scalar loop.
2068 PHINode *ResumeIndex = 0;
2069 LoopVectorizationLegality::InductionList::iterator I, E;
2070 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2071 // Set builder to point to last bypass block.
2072 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2073 for (I = List->begin(), E = List->end(); I != E; ++I) {
2074 PHINode *OrigPhi = I->first;
2075 LoopVectorizationLegality::InductionInfo II = I->second;
2077 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2078 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2079 MiddleBlock->getTerminator());
2080 // We might have extended the type of the induction variable but we need a
2081 // truncated version for the scalar loop.
2082 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2083 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2084 MiddleBlock->getTerminator()) : 0;
2086 Value *EndValue = 0;
2088 case LoopVectorizationLegality::IK_NoInduction:
2089 llvm_unreachable("Unknown induction");
2090 case LoopVectorizationLegality::IK_IntInduction: {
2091 // Handle the integer induction counter.
2092 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2094 // We have the canonical induction variable.
2095 if (OrigPhi == OldInduction) {
2096 // Create a truncated version of the resume value for the scalar loop,
2097 // we might have promoted the type to a larger width.
2099 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2100 // The new PHI merges the original incoming value, in case of a bypass,
2101 // or the value at the end of the vectorized loop.
2102 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2103 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2104 TruncResumeVal->addIncoming(EndValue, VecBody);
2106 // We know what the end value is.
2107 EndValue = IdxEndRoundDown;
2108 // We also know which PHI node holds it.
2109 ResumeIndex = ResumeVal;
2113 // Not the canonical induction variable - add the vector loop count to the
2115 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2116 II.StartValue->getType(),
2118 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2121 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2122 // Convert the CountRoundDown variable to the PHI size.
2123 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2124 II.StartValue->getType(),
2126 // Handle reverse integer induction counter.
2127 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2130 case LoopVectorizationLegality::IK_PtrInduction: {
2131 // For pointer induction variables, calculate the offset using
2133 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2137 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2138 // The value at the end of the loop for the reverse pointer is calculated
2139 // by creating a GEP with a negative index starting from the start value.
2140 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2141 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2143 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2149 // The new PHI merges the original incoming value, in case of a bypass,
2150 // or the value at the end of the vectorized loop.
2151 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2152 if (OrigPhi == OldInduction)
2153 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2155 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2157 ResumeVal->addIncoming(EndValue, VecBody);
2159 // Fix the scalar body counter (PHI node).
2160 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2161 // The old inductions phi node in the scalar body needs the truncated value.
2162 if (OrigPhi == OldInduction)
2163 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2165 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2168 // If we are generating a new induction variable then we also need to
2169 // generate the code that calculates the exit value. This value is not
2170 // simply the end of the counter because we may skip the vectorized body
2171 // in case of a runtime check.
2173 assert(!ResumeIndex && "Unexpected resume value found");
2174 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2175 MiddleBlock->getTerminator());
2176 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2177 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2178 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2181 // Make sure that we found the index where scalar loop needs to continue.
2182 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2183 "Invalid resume Index");
2185 // Add a check in the middle block to see if we have completed
2186 // all of the iterations in the first vector loop.
2187 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2188 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2189 ResumeIndex, "cmp.n",
2190 MiddleBlock->getTerminator());
2192 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2193 // Remove the old terminator.
2194 MiddleBlock->getTerminator()->eraseFromParent();
2196 // Create i+1 and fill the PHINode.
2197 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2198 Induction->addIncoming(StartIdx, VectorPH);
2199 Induction->addIncoming(NextIdx, VecBody);
2200 // Create the compare.
2201 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2202 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2204 // Now we have two terminators. Remove the old one from the block.
2205 VecBody->getTerminator()->eraseFromParent();
2207 // Get ready to start creating new instructions into the vectorized body.
2208 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2211 LoopVectorPreHeader = VectorPH;
2212 LoopScalarPreHeader = ScalarPH;
2213 LoopMiddleBlock = MiddleBlock;
2214 LoopExitBlock = ExitBlock;
2215 LoopVectorBody.push_back(VecBody);
2216 LoopScalarBody = OldBasicBlock;
2218 LoopVectorizeHints Hints(Lp, true);
2219 Hints.setAlreadyVectorized(Lp);
2222 /// This function returns the identity element (or neutral element) for
2223 /// the operation K.
2225 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2230 // Adding, Xoring, Oring zero to a number does not change it.
2231 return ConstantInt::get(Tp, 0);
2232 case RK_IntegerMult:
2233 // Multiplying a number by 1 does not change it.
2234 return ConstantInt::get(Tp, 1);
2236 // AND-ing a number with an all-1 value does not change it.
2237 return ConstantInt::get(Tp, -1, true);
2239 // Multiplying a number by 1 does not change it.
2240 return ConstantFP::get(Tp, 1.0L);
2242 // Adding zero to a number does not change it.
2243 return ConstantFP::get(Tp, 0.0L);
2245 llvm_unreachable("Unknown reduction kind");
2249 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2250 Intrinsic::ID ValidIntrinsicID) {
2251 if (I.getNumArgOperands() != 1 ||
2252 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2253 I.getType() != I.getArgOperand(0)->getType() ||
2254 !I.onlyReadsMemory())
2255 return Intrinsic::not_intrinsic;
2257 return ValidIntrinsicID;
2260 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2261 Intrinsic::ID ValidIntrinsicID) {
2262 if (I.getNumArgOperands() != 2 ||
2263 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2264 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2265 I.getType() != I.getArgOperand(0)->getType() ||
2266 I.getType() != I.getArgOperand(1)->getType() ||
2267 !I.onlyReadsMemory())
2268 return Intrinsic::not_intrinsic;
2270 return ValidIntrinsicID;
2274 static Intrinsic::ID
2275 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2276 // If we have an intrinsic call, check if it is trivially vectorizable.
2277 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2278 Intrinsic::ID ID = II->getIntrinsicID();
2279 if (isTriviallyVectorizable(ID) || ID == Intrinsic::lifetime_start ||
2280 ID == Intrinsic::lifetime_end)
2283 return Intrinsic::not_intrinsic;
2287 return Intrinsic::not_intrinsic;
2290 Function *F = CI->getCalledFunction();
2291 // We're going to make assumptions on the semantics of the functions, check
2292 // that the target knows that it's available in this environment and it does
2293 // not have local linkage.
2294 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2295 return Intrinsic::not_intrinsic;
2297 // Otherwise check if we have a call to a function that can be turned into a
2298 // vector intrinsic.
2305 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2309 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2313 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2315 case LibFunc::exp2f:
2316 case LibFunc::exp2l:
2317 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2321 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2322 case LibFunc::log10:
2323 case LibFunc::log10f:
2324 case LibFunc::log10l:
2325 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2327 case LibFunc::log2f:
2328 case LibFunc::log2l:
2329 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2331 case LibFunc::fabsf:
2332 case LibFunc::fabsl:
2333 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2334 case LibFunc::copysign:
2335 case LibFunc::copysignf:
2336 case LibFunc::copysignl:
2337 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2338 case LibFunc::floor:
2339 case LibFunc::floorf:
2340 case LibFunc::floorl:
2341 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2343 case LibFunc::ceilf:
2344 case LibFunc::ceill:
2345 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2346 case LibFunc::trunc:
2347 case LibFunc::truncf:
2348 case LibFunc::truncl:
2349 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2351 case LibFunc::rintf:
2352 case LibFunc::rintl:
2353 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2354 case LibFunc::nearbyint:
2355 case LibFunc::nearbyintf:
2356 case LibFunc::nearbyintl:
2357 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2358 case LibFunc::round:
2359 case LibFunc::roundf:
2360 case LibFunc::roundl:
2361 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2365 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2368 return Intrinsic::not_intrinsic;
2371 /// This function translates the reduction kind to an LLVM binary operator.
2373 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2375 case LoopVectorizationLegality::RK_IntegerAdd:
2376 return Instruction::Add;
2377 case LoopVectorizationLegality::RK_IntegerMult:
2378 return Instruction::Mul;
2379 case LoopVectorizationLegality::RK_IntegerOr:
2380 return Instruction::Or;
2381 case LoopVectorizationLegality::RK_IntegerAnd:
2382 return Instruction::And;
2383 case LoopVectorizationLegality::RK_IntegerXor:
2384 return Instruction::Xor;
2385 case LoopVectorizationLegality::RK_FloatMult:
2386 return Instruction::FMul;
2387 case LoopVectorizationLegality::RK_FloatAdd:
2388 return Instruction::FAdd;
2389 case LoopVectorizationLegality::RK_IntegerMinMax:
2390 return Instruction::ICmp;
2391 case LoopVectorizationLegality::RK_FloatMinMax:
2392 return Instruction::FCmp;
2394 llvm_unreachable("Unknown reduction operation");
2398 Value *createMinMaxOp(IRBuilder<> &Builder,
2399 LoopVectorizationLegality::MinMaxReductionKind RK,
2402 CmpInst::Predicate P = CmpInst::ICMP_NE;
2405 llvm_unreachable("Unknown min/max reduction kind");
2406 case LoopVectorizationLegality::MRK_UIntMin:
2407 P = CmpInst::ICMP_ULT;
2409 case LoopVectorizationLegality::MRK_UIntMax:
2410 P = CmpInst::ICMP_UGT;
2412 case LoopVectorizationLegality::MRK_SIntMin:
2413 P = CmpInst::ICMP_SLT;
2415 case LoopVectorizationLegality::MRK_SIntMax:
2416 P = CmpInst::ICMP_SGT;
2418 case LoopVectorizationLegality::MRK_FloatMin:
2419 P = CmpInst::FCMP_OLT;
2421 case LoopVectorizationLegality::MRK_FloatMax:
2422 P = CmpInst::FCMP_OGT;
2427 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2428 RK == LoopVectorizationLegality::MRK_FloatMax)
2429 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2431 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2433 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2438 struct CSEDenseMapInfo {
2439 static bool canHandle(Instruction *I) {
2440 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2441 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2443 static inline Instruction *getEmptyKey() {
2444 return DenseMapInfo<Instruction *>::getEmptyKey();
2446 static inline Instruction *getTombstoneKey() {
2447 return DenseMapInfo<Instruction *>::getTombstoneKey();
2449 static unsigned getHashValue(Instruction *I) {
2450 assert(canHandle(I) && "Unknown instruction!");
2451 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2452 I->value_op_end()));
2454 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2455 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2456 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2458 return LHS->isIdenticalTo(RHS);
2463 /// \brief Check whether this block is a predicated block.
2464 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2465 /// = ...; " blocks. We start with one vectorized basic block. For every
2466 /// conditional block we split this vectorized block. Therefore, every second
2467 /// block will be a predicated one.
2468 static bool isPredicatedBlock(unsigned BlockNum) {
2469 return BlockNum % 2;
2472 ///\brief Perform cse of induction variable instructions.
2473 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2474 // Perform simple cse.
2475 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2476 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2477 BasicBlock *BB = BBs[i];
2478 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2479 Instruction *In = I++;
2481 if (!CSEDenseMapInfo::canHandle(In))
2484 // Check if we can replace this instruction with any of the
2485 // visited instructions.
2486 if (Instruction *V = CSEMap.lookup(In)) {
2487 In->replaceAllUsesWith(V);
2488 In->eraseFromParent();
2491 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2492 // ...;" blocks for predicated stores. Every second block is a predicated
2494 if (isPredicatedBlock(i))
2502 /// \brief Adds a 'fast' flag to floating point operations.
2503 static Value *addFastMathFlag(Value *V) {
2504 if (isa<FPMathOperator>(V)){
2505 FastMathFlags Flags;
2506 Flags.setUnsafeAlgebra();
2507 cast<Instruction>(V)->setFastMathFlags(Flags);
2512 void InnerLoopVectorizer::vectorizeLoop() {
2513 //===------------------------------------------------===//
2515 // Notice: any optimization or new instruction that go
2516 // into the code below should be also be implemented in
2519 //===------------------------------------------------===//
2520 Constant *Zero = Builder.getInt32(0);
2522 // In order to support reduction variables we need to be able to vectorize
2523 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2524 // stages. First, we create a new vector PHI node with no incoming edges.
2525 // We use this value when we vectorize all of the instructions that use the
2526 // PHI. Next, after all of the instructions in the block are complete we
2527 // add the new incoming edges to the PHI. At this point all of the
2528 // instructions in the basic block are vectorized, so we can use them to
2529 // construct the PHI.
2530 PhiVector RdxPHIsToFix;
2532 // Scan the loop in a topological order to ensure that defs are vectorized
2534 LoopBlocksDFS DFS(OrigLoop);
2537 // Vectorize all of the blocks in the original loop.
2538 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2539 be = DFS.endRPO(); bb != be; ++bb)
2540 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2542 // At this point every instruction in the original loop is widened to
2543 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2544 // that we vectorized. The PHI nodes are currently empty because we did
2545 // not want to introduce cycles. Notice that the remaining PHI nodes
2546 // that we need to fix are reduction variables.
2548 // Create the 'reduced' values for each of the induction vars.
2549 // The reduced values are the vector values that we scalarize and combine
2550 // after the loop is finished.
2551 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2553 PHINode *RdxPhi = *it;
2554 assert(RdxPhi && "Unable to recover vectorized PHI");
2556 // Find the reduction variable descriptor.
2557 assert(Legal->getReductionVars()->count(RdxPhi) &&
2558 "Unable to find the reduction variable");
2559 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2560 (*Legal->getReductionVars())[RdxPhi];
2562 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2564 // We need to generate a reduction vector from the incoming scalar.
2565 // To do so, we need to generate the 'identity' vector and override
2566 // one of the elements with the incoming scalar reduction. We need
2567 // to do it in the vector-loop preheader.
2568 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2570 // This is the vector-clone of the value that leaves the loop.
2571 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2572 Type *VecTy = VectorExit[0]->getType();
2574 // Find the reduction identity variable. Zero for addition, or, xor,
2575 // one for multiplication, -1 for And.
2578 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2579 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2580 // MinMax reduction have the start value as their identify.
2582 VectorStart = Identity = RdxDesc.StartValue;
2584 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2589 // Handle other reduction kinds:
2591 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2592 VecTy->getScalarType());
2595 // This vector is the Identity vector where the first element is the
2596 // incoming scalar reduction.
2597 VectorStart = RdxDesc.StartValue;
2599 Identity = ConstantVector::getSplat(VF, Iden);
2601 // This vector is the Identity vector where the first element is the
2602 // incoming scalar reduction.
2603 VectorStart = Builder.CreateInsertElement(Identity,
2604 RdxDesc.StartValue, Zero);
2608 // Fix the vector-loop phi.
2609 // We created the induction variable so we know that the
2610 // preheader is the first entry.
2611 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2613 // Reductions do not have to start at zero. They can start with
2614 // any loop invariant values.
2615 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2616 BasicBlock *Latch = OrigLoop->getLoopLatch();
2617 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2618 VectorParts &Val = getVectorValue(LoopVal);
2619 for (unsigned part = 0; part < UF; ++part) {
2620 // Make sure to add the reduction stat value only to the
2621 // first unroll part.
2622 Value *StartVal = (part == 0) ? VectorStart : Identity;
2623 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2624 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2625 LoopVectorBody.back());
2628 // Before each round, move the insertion point right between
2629 // the PHIs and the values we are going to write.
2630 // This allows us to write both PHINodes and the extractelement
2632 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2634 VectorParts RdxParts;
2635 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2636 for (unsigned part = 0; part < UF; ++part) {
2637 // This PHINode contains the vectorized reduction variable, or
2638 // the initial value vector, if we bypass the vector loop.
2639 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2640 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2641 Value *StartVal = (part == 0) ? VectorStart : Identity;
2642 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2643 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2644 NewPhi->addIncoming(RdxExitVal[part],
2645 LoopVectorBody.back());
2646 RdxParts.push_back(NewPhi);
2649 // Reduce all of the unrolled parts into a single vector.
2650 Value *ReducedPartRdx = RdxParts[0];
2651 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2652 setDebugLocFromInst(Builder, ReducedPartRdx);
2653 for (unsigned part = 1; part < UF; ++part) {
2654 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2655 // Floating point operations had to be 'fast' to enable the reduction.
2656 ReducedPartRdx = addFastMathFlag(
2657 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2658 ReducedPartRdx, "bin.rdx"));
2660 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2661 ReducedPartRdx, RdxParts[part]);
2665 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2666 // and vector ops, reducing the set of values being computed by half each
2668 assert(isPowerOf2_32(VF) &&
2669 "Reduction emission only supported for pow2 vectors!");
2670 Value *TmpVec = ReducedPartRdx;
2671 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2672 for (unsigned i = VF; i != 1; i >>= 1) {
2673 // Move the upper half of the vector to the lower half.
2674 for (unsigned j = 0; j != i/2; ++j)
2675 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2677 // Fill the rest of the mask with undef.
2678 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2679 UndefValue::get(Builder.getInt32Ty()));
2682 Builder.CreateShuffleVector(TmpVec,
2683 UndefValue::get(TmpVec->getType()),
2684 ConstantVector::get(ShuffleMask),
2687 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2688 // Floating point operations had to be 'fast' to enable the reduction.
2689 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2690 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2692 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2695 // The result is in the first element of the vector.
2696 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2697 Builder.getInt32(0));
2700 // Now, we need to fix the users of the reduction variable
2701 // inside and outside of the scalar remainder loop.
2702 // We know that the loop is in LCSSA form. We need to update the
2703 // PHI nodes in the exit blocks.
2704 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2705 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2706 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2707 if (!LCSSAPhi) break;
2709 // All PHINodes need to have a single entry edge, or two if
2710 // we already fixed them.
2711 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2713 // We found our reduction value exit-PHI. Update it with the
2714 // incoming bypass edge.
2715 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2716 // Add an edge coming from the bypass.
2717 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2720 }// end of the LCSSA phi scan.
2722 // Fix the scalar loop reduction variable with the incoming reduction sum
2723 // from the vector body and from the backedge value.
2724 int IncomingEdgeBlockIdx =
2725 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2726 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2727 // Pick the other block.
2728 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2729 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2730 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2731 }// end of for each redux variable.
2735 // Remove redundant induction instructions.
2736 cse(LoopVectorBody);
2739 void InnerLoopVectorizer::fixLCSSAPHIs() {
2740 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2741 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2742 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2743 if (!LCSSAPhi) break;
2744 if (LCSSAPhi->getNumIncomingValues() == 1)
2745 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2750 InnerLoopVectorizer::VectorParts
2751 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2752 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2755 // Look for cached value.
2756 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2757 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2758 if (ECEntryIt != MaskCache.end())
2759 return ECEntryIt->second;
2761 VectorParts SrcMask = createBlockInMask(Src);
2763 // The terminator has to be a branch inst!
2764 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2765 assert(BI && "Unexpected terminator found");
2767 if (BI->isConditional()) {
2768 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2770 if (BI->getSuccessor(0) != Dst)
2771 for (unsigned part = 0; part < UF; ++part)
2772 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2774 for (unsigned part = 0; part < UF; ++part)
2775 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2777 MaskCache[Edge] = EdgeMask;
2781 MaskCache[Edge] = SrcMask;
2785 InnerLoopVectorizer::VectorParts
2786 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2787 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2789 // Loop incoming mask is all-one.
2790 if (OrigLoop->getHeader() == BB) {
2791 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2792 return getVectorValue(C);
2795 // This is the block mask. We OR all incoming edges, and with zero.
2796 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2797 VectorParts BlockMask = getVectorValue(Zero);
2800 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2801 VectorParts EM = createEdgeMask(*it, BB);
2802 for (unsigned part = 0; part < UF; ++part)
2803 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2809 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2810 InnerLoopVectorizer::VectorParts &Entry,
2811 unsigned UF, unsigned VF, PhiVector *PV) {
2812 PHINode* P = cast<PHINode>(PN);
2813 // Handle reduction variables:
2814 if (Legal->getReductionVars()->count(P)) {
2815 for (unsigned part = 0; part < UF; ++part) {
2816 // This is phase one of vectorizing PHIs.
2817 Type *VecTy = (VF == 1) ? PN->getType() :
2818 VectorType::get(PN->getType(), VF);
2819 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2820 LoopVectorBody.back()-> getFirstInsertionPt());
2826 setDebugLocFromInst(Builder, P);
2827 // Check for PHI nodes that are lowered to vector selects.
2828 if (P->getParent() != OrigLoop->getHeader()) {
2829 // We know that all PHIs in non-header blocks are converted into
2830 // selects, so we don't have to worry about the insertion order and we
2831 // can just use the builder.
2832 // At this point we generate the predication tree. There may be
2833 // duplications since this is a simple recursive scan, but future
2834 // optimizations will clean it up.
2836 unsigned NumIncoming = P->getNumIncomingValues();
2838 // Generate a sequence of selects of the form:
2839 // SELECT(Mask3, In3,
2840 // SELECT(Mask2, In2,
2842 for (unsigned In = 0; In < NumIncoming; In++) {
2843 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2845 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2847 for (unsigned part = 0; part < UF; ++part) {
2848 // We might have single edge PHIs (blocks) - use an identity
2849 // 'select' for the first PHI operand.
2851 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2854 // Select between the current value and the previous incoming edge
2855 // based on the incoming mask.
2856 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2857 Entry[part], "predphi");
2863 // This PHINode must be an induction variable.
2864 // Make sure that we know about it.
2865 assert(Legal->getInductionVars()->count(P) &&
2866 "Not an induction variable");
2868 LoopVectorizationLegality::InductionInfo II =
2869 Legal->getInductionVars()->lookup(P);
2872 case LoopVectorizationLegality::IK_NoInduction:
2873 llvm_unreachable("Unknown induction");
2874 case LoopVectorizationLegality::IK_IntInduction: {
2875 assert(P->getType() == II.StartValue->getType() && "Types must match");
2876 Type *PhiTy = P->getType();
2878 if (P == OldInduction) {
2879 // Handle the canonical induction variable. We might have had to
2881 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2883 // Handle other induction variables that are now based on the
2885 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2887 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2888 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2891 Broadcasted = getBroadcastInstrs(Broadcasted);
2892 // After broadcasting the induction variable we need to make the vector
2893 // consecutive by adding 0, 1, 2, etc.
2894 for (unsigned part = 0; part < UF; ++part)
2895 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2898 case LoopVectorizationLegality::IK_ReverseIntInduction:
2899 case LoopVectorizationLegality::IK_PtrInduction:
2900 case LoopVectorizationLegality::IK_ReversePtrInduction:
2901 // Handle reverse integer and pointer inductions.
2902 Value *StartIdx = ExtendedIdx;
2903 // This is the normalized GEP that starts counting at zero.
2904 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2907 // Handle the reverse integer induction variable case.
2908 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2909 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2910 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2912 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2915 // This is a new value so do not hoist it out.
2916 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2917 // After broadcasting the induction variable we need to make the
2918 // vector consecutive by adding ... -3, -2, -1, 0.
2919 for (unsigned part = 0; part < UF; ++part)
2920 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2925 // Handle the pointer induction variable case.
2926 assert(P->getType()->isPointerTy() && "Unexpected type.");
2928 // Is this a reverse induction ptr or a consecutive induction ptr.
2929 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2932 // This is the vector of results. Notice that we don't generate
2933 // vector geps because scalar geps result in better code.
2934 for (unsigned part = 0; part < UF; ++part) {
2936 int EltIndex = (part) * (Reverse ? -1 : 1);
2937 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2940 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2942 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2944 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2946 Entry[part] = SclrGep;
2950 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2951 for (unsigned int i = 0; i < VF; ++i) {
2952 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2953 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2956 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2958 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2960 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2962 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2963 Builder.getInt32(i),
2966 Entry[part] = VecVal;
2972 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2973 // For each instruction in the old loop.
2974 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2975 VectorParts &Entry = WidenMap.get(it);
2976 switch (it->getOpcode()) {
2977 case Instruction::Br:
2978 // Nothing to do for PHIs and BR, since we already took care of the
2979 // loop control flow instructions.
2981 case Instruction::PHI:{
2982 // Vectorize PHINodes.
2983 widenPHIInstruction(it, Entry, UF, VF, PV);
2987 case Instruction::Add:
2988 case Instruction::FAdd:
2989 case Instruction::Sub:
2990 case Instruction::FSub:
2991 case Instruction::Mul:
2992 case Instruction::FMul:
2993 case Instruction::UDiv:
2994 case Instruction::SDiv:
2995 case Instruction::FDiv:
2996 case Instruction::URem:
2997 case Instruction::SRem:
2998 case Instruction::FRem:
2999 case Instruction::Shl:
3000 case Instruction::LShr:
3001 case Instruction::AShr:
3002 case Instruction::And:
3003 case Instruction::Or:
3004 case Instruction::Xor: {
3005 // Just widen binops.
3006 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3007 setDebugLocFromInst(Builder, BinOp);
3008 VectorParts &A = getVectorValue(it->getOperand(0));
3009 VectorParts &B = getVectorValue(it->getOperand(1));
3011 // Use this vector value for all users of the original instruction.
3012 for (unsigned Part = 0; Part < UF; ++Part) {
3013 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3015 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3016 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3017 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3018 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3019 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3021 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3022 VecOp->setIsExact(BinOp->isExact());
3024 // Copy the fast-math flags.
3025 if (VecOp && isa<FPMathOperator>(V))
3026 VecOp->setFastMathFlags(it->getFastMathFlags());
3032 case Instruction::Select: {
3034 // If the selector is loop invariant we can create a select
3035 // instruction with a scalar condition. Otherwise, use vector-select.
3036 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3038 setDebugLocFromInst(Builder, it);
3040 // The condition can be loop invariant but still defined inside the
3041 // loop. This means that we can't just use the original 'cond' value.
3042 // We have to take the 'vectorized' value and pick the first lane.
3043 // Instcombine will make this a no-op.
3044 VectorParts &Cond = getVectorValue(it->getOperand(0));
3045 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3046 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3048 Value *ScalarCond = (VF == 1) ? Cond[0] :
3049 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3051 for (unsigned Part = 0; Part < UF; ++Part) {
3052 Entry[Part] = Builder.CreateSelect(
3053 InvariantCond ? ScalarCond : Cond[Part],
3060 case Instruction::ICmp:
3061 case Instruction::FCmp: {
3062 // Widen compares. Generate vector compares.
3063 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3064 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3065 setDebugLocFromInst(Builder, it);
3066 VectorParts &A = getVectorValue(it->getOperand(0));
3067 VectorParts &B = getVectorValue(it->getOperand(1));
3068 for (unsigned Part = 0; Part < UF; ++Part) {
3071 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3073 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3079 case Instruction::Store:
3080 case Instruction::Load:
3081 vectorizeMemoryInstruction(it);
3083 case Instruction::ZExt:
3084 case Instruction::SExt:
3085 case Instruction::FPToUI:
3086 case Instruction::FPToSI:
3087 case Instruction::FPExt:
3088 case Instruction::PtrToInt:
3089 case Instruction::IntToPtr:
3090 case Instruction::SIToFP:
3091 case Instruction::UIToFP:
3092 case Instruction::Trunc:
3093 case Instruction::FPTrunc:
3094 case Instruction::BitCast: {
3095 CastInst *CI = dyn_cast<CastInst>(it);
3096 setDebugLocFromInst(Builder, it);
3097 /// Optimize the special case where the source is the induction
3098 /// variable. Notice that we can only optimize the 'trunc' case
3099 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3100 /// c. other casts depend on pointer size.
3101 if (CI->getOperand(0) == OldInduction &&
3102 it->getOpcode() == Instruction::Trunc) {
3103 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3105 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3106 for (unsigned Part = 0; Part < UF; ++Part)
3107 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3110 /// Vectorize casts.
3111 Type *DestTy = (VF == 1) ? CI->getType() :
3112 VectorType::get(CI->getType(), VF);
3114 VectorParts &A = getVectorValue(it->getOperand(0));
3115 for (unsigned Part = 0; Part < UF; ++Part)
3116 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3120 case Instruction::Call: {
3121 // Ignore dbg intrinsics.
3122 if (isa<DbgInfoIntrinsic>(it))
3124 setDebugLocFromInst(Builder, it);
3126 Module *M = BB->getParent()->getParent();
3127 CallInst *CI = cast<CallInst>(it);
3128 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3129 assert(ID && "Not an intrinsic call!");
3131 case Intrinsic::lifetime_end:
3132 case Intrinsic::lifetime_start:
3133 scalarizeInstruction(it);
3136 for (unsigned Part = 0; Part < UF; ++Part) {
3137 SmallVector<Value *, 4> Args;
3138 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3139 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3140 Args.push_back(Arg[Part]);
3142 Type *Tys[] = {CI->getType()};
3144 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3146 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3147 Entry[Part] = Builder.CreateCall(F, Args);
3155 // All other instructions are unsupported. Scalarize them.
3156 scalarizeInstruction(it);
3159 }// end of for_each instr.
3162 void InnerLoopVectorizer::updateAnalysis() {
3163 // Forget the original basic block.
3164 SE->forgetLoop(OrigLoop);
3166 // Update the dominator tree information.
3167 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3168 "Entry does not dominate exit.");
3170 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3171 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3172 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3174 // Due to if predication of stores we might create a sequence of "if(pred)
3175 // a[i] = ...; " blocks.
3176 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3178 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3179 else if (isPredicatedBlock(i)) {
3180 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3182 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3186 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3187 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3188 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3189 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3191 DEBUG(DT->verifyDomTree());
3194 /// \brief Check whether it is safe to if-convert this phi node.
3196 /// Phi nodes with constant expressions that can trap are not safe to if
3198 static bool canIfConvertPHINodes(BasicBlock *BB) {
3199 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3200 PHINode *Phi = dyn_cast<PHINode>(I);
3203 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3204 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3211 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3212 if (!EnableIfConversion)
3215 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3217 // A list of pointers that we can safely read and write to.
3218 SmallPtrSet<Value *, 8> SafePointes;
3220 // Collect safe addresses.
3221 for (Loop::block_iterator BI = TheLoop->block_begin(),
3222 BE = TheLoop->block_end(); BI != BE; ++BI) {
3223 BasicBlock *BB = *BI;
3225 if (blockNeedsPredication(BB))
3228 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3229 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3230 SafePointes.insert(LI->getPointerOperand());
3231 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3232 SafePointes.insert(SI->getPointerOperand());
3236 // Collect the blocks that need predication.
3237 BasicBlock *Header = TheLoop->getHeader();
3238 for (Loop::block_iterator BI = TheLoop->block_begin(),
3239 BE = TheLoop->block_end(); BI != BE; ++BI) {
3240 BasicBlock *BB = *BI;
3242 // We don't support switch statements inside loops.
3243 if (!isa<BranchInst>(BB->getTerminator()))
3246 // We must be able to predicate all blocks that need to be predicated.
3247 if (blockNeedsPredication(BB)) {
3248 if (!blockCanBePredicated(BB, SafePointes))
3250 } else if (BB != Header && !canIfConvertPHINodes(BB))
3255 // We can if-convert this loop.
3259 bool LoopVectorizationLegality::canVectorize() {
3260 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3261 // be canonicalized.
3262 if (!TheLoop->getLoopPreheader())
3265 // We can only vectorize innermost loops.
3266 if (TheLoop->getSubLoopsVector().size())
3269 // We must have a single backedge.
3270 if (TheLoop->getNumBackEdges() != 1)
3273 // We must have a single exiting block.
3274 if (!TheLoop->getExitingBlock())
3277 // We need to have a loop header.
3278 DEBUG(dbgs() << "LV: Found a loop: " <<
3279 TheLoop->getHeader()->getName() << '\n');
3281 // Check if we can if-convert non-single-bb loops.
3282 unsigned NumBlocks = TheLoop->getNumBlocks();
3283 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3284 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3288 // ScalarEvolution needs to be able to find the exit count.
3289 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3290 if (ExitCount == SE->getCouldNotCompute()) {
3291 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3295 // Do not loop-vectorize loops with a tiny trip count.
3296 BasicBlock *Latch = TheLoop->getLoopLatch();
3297 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
3298 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
3299 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
3300 "This loop is not worth vectorizing.\n");
3304 // Check if we can vectorize the instructions and CFG in this loop.
3305 if (!canVectorizeInstrs()) {
3306 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3310 // Go over each instruction and look at memory deps.
3311 if (!canVectorizeMemory()) {
3312 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3316 // Collect all of the variables that remain uniform after vectorization.
3317 collectLoopUniforms();
3319 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3320 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3323 // Okay! We can vectorize. At this point we don't have any other mem analysis
3324 // which may limit our maximum vectorization factor, so just return true with
3329 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3330 if (Ty->isPointerTy())
3331 return DL.getIntPtrType(Ty);
3333 // It is possible that char's or short's overflow when we ask for the loop's
3334 // trip count, work around this by changing the type size.
3335 if (Ty->getScalarSizeInBits() < 32)
3336 return Type::getInt32Ty(Ty->getContext());
3341 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3342 Ty0 = convertPointerToIntegerType(DL, Ty0);
3343 Ty1 = convertPointerToIntegerType(DL, Ty1);
3344 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3349 /// \brief Check that the instruction has outside loop users and is not an
3350 /// identified reduction variable.
3351 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3352 SmallPtrSet<Value *, 4> &Reductions) {
3353 // Reduction instructions are allowed to have exit users. All other
3354 // instructions must not have external users.
3355 if (!Reductions.count(Inst))
3356 //Check that all of the users of the loop are inside the BB.
3357 for (User *U : Inst->users()) {
3358 Instruction *UI = cast<Instruction>(U);
3359 // This user may be a reduction exit value.
3360 if (!TheLoop->contains(UI)) {
3361 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3368 bool LoopVectorizationLegality::canVectorizeInstrs() {
3369 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3370 BasicBlock *Header = TheLoop->getHeader();
3372 // Look for the attribute signaling the absence of NaNs.
3373 Function &F = *Header->getParent();
3374 if (F.hasFnAttribute("no-nans-fp-math"))
3375 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3376 AttributeSet::FunctionIndex,
3377 "no-nans-fp-math").getValueAsString() == "true";
3379 // For each block in the loop.
3380 for (Loop::block_iterator bb = TheLoop->block_begin(),
3381 be = TheLoop->block_end(); bb != be; ++bb) {
3383 // Scan the instructions in the block and look for hazards.
3384 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3387 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3388 Type *PhiTy = Phi->getType();
3389 // Check that this PHI type is allowed.
3390 if (!PhiTy->isIntegerTy() &&
3391 !PhiTy->isFloatingPointTy() &&
3392 !PhiTy->isPointerTy()) {
3393 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3397 // If this PHINode is not in the header block, then we know that we
3398 // can convert it to select during if-conversion. No need to check if
3399 // the PHIs in this block are induction or reduction variables.
3400 if (*bb != Header) {
3401 // Check that this instruction has no outside users or is an
3402 // identified reduction value with an outside user.
3403 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3408 // We only allow if-converted PHIs with more than two incoming values.
3409 if (Phi->getNumIncomingValues() != 2) {
3410 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3414 // This is the value coming from the preheader.
3415 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3416 // Check if this is an induction variable.
3417 InductionKind IK = isInductionVariable(Phi);
3419 if (IK_NoInduction != IK) {
3420 // Get the widest type.
3422 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3424 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3426 // Int inductions are special because we only allow one IV.
3427 if (IK == IK_IntInduction) {
3428 // Use the phi node with the widest type as induction. Use the last
3429 // one if there are multiple (no good reason for doing this other
3430 // than it is expedient).
3431 if (!Induction || PhiTy == WidestIndTy)
3435 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3436 Inductions[Phi] = InductionInfo(StartValue, IK);
3438 // Until we explicitly handle the case of an induction variable with
3439 // an outside loop user we have to give up vectorizing this loop.
3440 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3446 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3447 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3450 if (AddReductionVar(Phi, RK_IntegerMult)) {
3451 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3454 if (AddReductionVar(Phi, RK_IntegerOr)) {
3455 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3458 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3459 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3462 if (AddReductionVar(Phi, RK_IntegerXor)) {
3463 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3466 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3467 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3470 if (AddReductionVar(Phi, RK_FloatMult)) {
3471 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3474 if (AddReductionVar(Phi, RK_FloatAdd)) {
3475 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3478 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3479 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3484 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3486 }// end of PHI handling
3488 // We still don't handle functions. However, we can ignore dbg intrinsic
3489 // calls and we do handle certain intrinsic and libm functions.
3490 CallInst *CI = dyn_cast<CallInst>(it);
3491 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3492 DEBUG(dbgs() << "LV: Found a call site.\n");
3496 // Check that the instruction return type is vectorizable.
3497 // Also, we can't vectorize extractelement instructions.
3498 if ((!VectorType::isValidElementType(it->getType()) &&
3499 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3500 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3504 // Check that the stored type is vectorizable.
3505 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3506 Type *T = ST->getValueOperand()->getType();
3507 if (!VectorType::isValidElementType(T))
3509 if (EnableMemAccessVersioning)
3510 collectStridedAcccess(ST);
3513 if (EnableMemAccessVersioning)
3514 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3515 collectStridedAcccess(LI);
3517 // Reduction instructions are allowed to have exit users.
3518 // All other instructions must not have external users.
3519 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3527 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3528 if (Inductions.empty())
3535 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3536 /// return the induction operand of the gep pointer.
3537 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3538 const DataLayout *DL, Loop *Lp) {
3539 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3543 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3545 // Check that all of the gep indices are uniform except for our induction
3547 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3548 if (i != InductionOperand &&
3549 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3551 return GEP->getOperand(InductionOperand);
3554 ///\brief Look for a cast use of the passed value.
3555 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3556 Value *UniqueCast = 0;
3557 for (User *U : Ptr->users()) {
3558 CastInst *CI = dyn_cast<CastInst>(U);
3559 if (CI && CI->getType() == Ty) {
3569 ///\brief Get the stride of a pointer access in a loop.
3570 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3571 /// pointer to the Value, or null otherwise.
3572 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3573 const DataLayout *DL, Loop *Lp) {
3574 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3575 if (!PtrTy || PtrTy->isAggregateType())
3578 // Try to remove a gep instruction to make the pointer (actually index at this
3579 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3580 // pointer, otherwise, we are analyzing the index.
3581 Value *OrigPtr = Ptr;
3583 // The size of the pointer access.
3584 int64_t PtrAccessSize = 1;
3586 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3587 const SCEV *V = SE->getSCEV(Ptr);
3591 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3592 V = C->getOperand();
3594 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3598 V = S->getStepRecurrence(*SE);
3602 // Strip off the size of access multiplication if we are still analyzing the
3604 if (OrigPtr == Ptr) {
3605 DL->getTypeAllocSize(PtrTy->getElementType());
3606 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3607 if (M->getOperand(0)->getSCEVType() != scConstant)
3610 const APInt &APStepVal =
3611 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3613 // Huge step value - give up.
3614 if (APStepVal.getBitWidth() > 64)
3617 int64_t StepVal = APStepVal.getSExtValue();
3618 if (PtrAccessSize != StepVal)
3620 V = M->getOperand(1);
3625 Type *StripedOffRecurrenceCast = 0;
3626 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3627 StripedOffRecurrenceCast = C->getType();
3628 V = C->getOperand();
3631 // Look for the loop invariant symbolic value.
3632 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3636 Value *Stride = U->getValue();
3637 if (!Lp->isLoopInvariant(Stride))
3640 // If we have stripped off the recurrence cast we have to make sure that we
3641 // return the value that is used in this loop so that we can replace it later.
3642 if (StripedOffRecurrenceCast)
3643 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3648 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3650 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3651 Ptr = LI->getPointerOperand();
3652 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3653 Ptr = SI->getPointerOperand();
3657 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3661 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3662 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3663 Strides[Ptr] = Stride;
3664 StrideSet.insert(Stride);
3667 void LoopVectorizationLegality::collectLoopUniforms() {
3668 // We now know that the loop is vectorizable!
3669 // Collect variables that will remain uniform after vectorization.
3670 std::vector<Value*> Worklist;
3671 BasicBlock *Latch = TheLoop->getLoopLatch();
3673 // Start with the conditional branch and walk up the block.
3674 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3676 // Also add all consecutive pointer values; these values will be uniform
3677 // after vectorization (and subsequent cleanup) and, until revectorization is
3678 // supported, all dependencies must also be uniform.
3679 for (Loop::block_iterator B = TheLoop->block_begin(),
3680 BE = TheLoop->block_end(); B != BE; ++B)
3681 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3683 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3684 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3686 while (Worklist.size()) {
3687 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3688 Worklist.pop_back();
3690 // Look at instructions inside this loop.
3691 // Stop when reaching PHI nodes.
3692 // TODO: we need to follow values all over the loop, not only in this block.
3693 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3696 // This is a known uniform.
3699 // Insert all operands.
3700 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3705 /// \brief Analyses memory accesses in a loop.
3707 /// Checks whether run time pointer checks are needed and builds sets for data
3708 /// dependence checking.
3709 class AccessAnalysis {
3711 /// \brief Read or write access location.
3712 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3713 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3715 /// \brief Set of potential dependent memory accesses.
3716 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3718 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3719 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3720 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3722 /// \brief Register a load and whether it is only read from.
3723 void addLoad(Value *Ptr, bool IsReadOnly) {
3724 Accesses.insert(MemAccessInfo(Ptr, false));
3726 ReadOnlyPtr.insert(Ptr);
3729 /// \brief Register a store.
3730 void addStore(Value *Ptr) {
3731 Accesses.insert(MemAccessInfo(Ptr, true));
3734 /// \brief Check whether we can check the pointers at runtime for
3735 /// non-intersection.
3736 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3737 unsigned &NumComparisons, ScalarEvolution *SE,
3738 Loop *TheLoop, ValueToValueMap &Strides,
3739 bool ShouldCheckStride = false);
3741 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3742 /// and builds sets of dependent accesses.
3743 void buildDependenceSets() {
3744 // Process read-write pointers first.
3745 processMemAccesses(false);
3746 // Next, process read pointers.
3747 processMemAccesses(true);
3750 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3752 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3753 void resetDepChecks() { CheckDeps.clear(); }
3755 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3758 typedef SetVector<MemAccessInfo> PtrAccessSet;
3759 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3761 /// \brief Go over all memory access or only the deferred ones if
3762 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3763 /// and build sets of dependency check candidates.
3764 void processMemAccesses(bool UseDeferred);
3766 /// Set of all accesses.
3767 PtrAccessSet Accesses;
3769 /// Set of access to check after all writes have been processed.
3770 PtrAccessSet DeferredAccesses;
3772 /// Map of pointers to last access encountered.
3773 UnderlyingObjToAccessMap ObjToLastAccess;
3775 /// Set of accesses that need a further dependence check.
3776 MemAccessInfoSet CheckDeps;
3778 /// Set of pointers that are read only.
3779 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3781 /// Set of underlying objects already written to.
3782 SmallPtrSet<Value*, 16> WriteObjects;
3784 const DataLayout *DL;
3786 /// Sets of potentially dependent accesses - members of one set share an
3787 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3788 /// dependence check.
3789 DepCandidates &DepCands;
3791 bool AreAllWritesIdentified;
3792 bool AreAllReadsIdentified;
3793 bool IsRTCheckNeeded;
3796 } // end anonymous namespace
3798 /// \brief Check whether a pointer can participate in a runtime bounds check.
3799 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3801 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3802 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3806 return AR->isAffine();
3809 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3810 /// the address space.
3811 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3812 const Loop *Lp, ValueToValueMap &StridesMap);
3814 bool AccessAnalysis::canCheckPtrAtRT(
3815 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3816 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3817 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3818 // Find pointers with computable bounds. We are going to use this information
3819 // to place a runtime bound check.
3820 unsigned NumReadPtrChecks = 0;
3821 unsigned NumWritePtrChecks = 0;
3822 bool CanDoRT = true;
3824 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3825 // We assign consecutive id to access from different dependence sets.
3826 // Accesses within the same set don't need a runtime check.
3827 unsigned RunningDepId = 1;
3828 DenseMap<Value *, unsigned> DepSetId;
3830 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3832 const MemAccessInfo &Access = *AI;
3833 Value *Ptr = Access.getPointer();
3834 bool IsWrite = Access.getInt();
3836 // Just add write checks if we have both.
3837 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3841 ++NumWritePtrChecks;
3845 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3846 // When we run after a failing dependency check we have to make sure we
3847 // don't have wrapping pointers.
3848 (!ShouldCheckStride ||
3849 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3850 // The id of the dependence set.
3853 if (IsDepCheckNeeded) {
3854 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3855 unsigned &LeaderId = DepSetId[Leader];
3857 LeaderId = RunningDepId++;
3860 // Each access has its own dependence set.
3861 DepId = RunningDepId++;
3863 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3865 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3871 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3872 NumComparisons = 0; // Only one dependence set.
3874 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3875 NumWritePtrChecks - 1));
3878 // If the pointers that we would use for the bounds comparison have different
3879 // address spaces, assume the values aren't directly comparable, so we can't
3880 // use them for the runtime check. We also have to assume they could
3881 // overlap. In the future there should be metadata for whether address spaces
3883 unsigned NumPointers = RtCheck.Pointers.size();
3884 for (unsigned i = 0; i < NumPointers; ++i) {
3885 for (unsigned j = i + 1; j < NumPointers; ++j) {
3886 // Only need to check pointers between two different dependency sets.
3887 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3890 Value *PtrI = RtCheck.Pointers[i];
3891 Value *PtrJ = RtCheck.Pointers[j];
3893 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3894 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3896 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3897 " different address spaces\n");
3906 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3907 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3910 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3911 // We process the set twice: first we process read-write pointers, last we
3912 // process read-only pointers. This allows us to skip dependence tests for
3913 // read-only pointers.
3915 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3916 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3917 const MemAccessInfo &Access = *AI;
3918 Value *Ptr = Access.getPointer();
3919 bool IsWrite = Access.getInt();
3921 DepCands.insert(Access);
3923 // Memorize read-only pointers for later processing and skip them in the
3924 // first round (they need to be checked after we have seen all write
3925 // pointers). Note: we also mark pointer that are not consecutive as
3926 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3927 // second check for "!IsWrite".
3928 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3929 if (!UseDeferred && IsReadOnlyPtr) {
3930 DeferredAccesses.insert(Access);
3934 bool NeedDepCheck = false;
3935 // Check whether there is the possibility of dependency because of
3936 // underlying objects being the same.
3937 typedef SmallVector<Value*, 16> ValueVector;
3938 ValueVector TempObjects;
3939 GetUnderlyingObjects(Ptr, TempObjects, DL);
3940 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3942 Value *UnderlyingObj = *UI;
3944 // If this is a write then it needs to be an identified object. If this a
3945 // read and all writes (so far) are identified function scope objects we
3946 // don't need an identified underlying object but only an Argument (the
3947 // next write is going to invalidate this assumption if it is
3949 // This is a micro-optimization for the case where all writes are
3950 // identified and we have one argument pointer.
3951 // Otherwise, we do need a runtime check.
3952 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3953 (!IsWrite && (!AreAllWritesIdentified ||
3954 !isa<Argument>(UnderlyingObj)) &&
3955 !isIdentifiedObject(UnderlyingObj))) {
3956 DEBUG(dbgs() << "LV: Found an unidentified " <<
3957 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3959 IsRTCheckNeeded = (IsRTCheckNeeded ||
3960 !isIdentifiedObject(UnderlyingObj) ||
3961 !AreAllReadsIdentified);
3964 AreAllWritesIdentified = false;
3966 AreAllReadsIdentified = false;
3969 // If this is a write - check other reads and writes for conflicts. If
3970 // this is a read only check other writes for conflicts (but only if there
3971 // is no other write to the ptr - this is an optimization to catch "a[i] =
3972 // a[i] + " without having to do a dependence check).
3973 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3974 NeedDepCheck = true;
3977 WriteObjects.insert(UnderlyingObj);
3979 // Create sets of pointers connected by shared underlying objects.
3980 UnderlyingObjToAccessMap::iterator Prev =
3981 ObjToLastAccess.find(UnderlyingObj);
3982 if (Prev != ObjToLastAccess.end())
3983 DepCands.unionSets(Access, Prev->second);
3985 ObjToLastAccess[UnderlyingObj] = Access;
3989 CheckDeps.insert(Access);
3994 /// \brief Checks memory dependences among accesses to the same underlying
3995 /// object to determine whether there vectorization is legal or not (and at
3996 /// which vectorization factor).
3998 /// This class works under the assumption that we already checked that memory
3999 /// locations with different underlying pointers are "must-not alias".
4000 /// We use the ScalarEvolution framework to symbolically evalutate access
4001 /// functions pairs. Since we currently don't restructure the loop we can rely
4002 /// on the program order of memory accesses to determine their safety.
4003 /// At the moment we will only deem accesses as safe for:
4004 /// * A negative constant distance assuming program order.
4006 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4007 /// a[i] = tmp; y = a[i];
4009 /// The latter case is safe because later checks guarantuee that there can't
4010 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4011 /// the same variable: a header phi can only be an induction or a reduction, a
4012 /// reduction can't have a memory sink, an induction can't have a memory
4013 /// source). This is important and must not be violated (or we have to
4014 /// resort to checking for cycles through memory).
4016 /// * A positive constant distance assuming program order that is bigger
4017 /// than the biggest memory access.
4019 /// tmp = a[i] OR b[i] = x
4020 /// a[i+2] = tmp y = b[i+2];
4022 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4024 /// * Zero distances and all accesses have the same size.
4026 class MemoryDepChecker {
4028 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4029 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4031 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4032 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4033 ShouldRetryWithRuntimeCheck(false) {}
4035 /// \brief Register the location (instructions are given increasing numbers)
4036 /// of a write access.
4037 void addAccess(StoreInst *SI) {
4038 Value *Ptr = SI->getPointerOperand();
4039 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4040 InstMap.push_back(SI);
4044 /// \brief Register the location (instructions are given increasing numbers)
4045 /// of a write access.
4046 void addAccess(LoadInst *LI) {
4047 Value *Ptr = LI->getPointerOperand();
4048 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4049 InstMap.push_back(LI);
4053 /// \brief Check whether the dependencies between the accesses are safe.
4055 /// Only checks sets with elements in \p CheckDeps.
4056 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4057 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4059 /// \brief The maximum number of bytes of a vector register we can vectorize
4060 /// the accesses safely with.
4061 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4063 /// \brief In same cases when the dependency check fails we can still
4064 /// vectorize the loop with a dynamic array access check.
4065 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4068 ScalarEvolution *SE;
4069 const DataLayout *DL;
4070 const Loop *InnermostLoop;
4072 /// \brief Maps access locations (ptr, read/write) to program order.
4073 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4075 /// \brief Memory access instructions in program order.
4076 SmallVector<Instruction *, 16> InstMap;
4078 /// \brief The program order index to be used for the next instruction.
4081 // We can access this many bytes in parallel safely.
4082 unsigned MaxSafeDepDistBytes;
4084 /// \brief If we see a non-constant dependence distance we can still try to
4085 /// vectorize this loop with runtime checks.
4086 bool ShouldRetryWithRuntimeCheck;
4088 /// \brief Check whether there is a plausible dependence between the two
4091 /// Access \p A must happen before \p B in program order. The two indices
4092 /// identify the index into the program order map.
4094 /// This function checks whether there is a plausible dependence (or the
4095 /// absence of such can't be proved) between the two accesses. If there is a
4096 /// plausible dependence but the dependence distance is bigger than one
4097 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4098 /// distance is smaller than any other distance encountered so far).
4099 /// Otherwise, this function returns true signaling a possible dependence.
4100 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4101 const MemAccessInfo &B, unsigned BIdx,
4102 ValueToValueMap &Strides);
4104 /// \brief Check whether the data dependence could prevent store-load
4106 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4109 } // end anonymous namespace
4111 static bool isInBoundsGep(Value *Ptr) {
4112 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4113 return GEP->isInBounds();
4117 /// \brief Check whether the access through \p Ptr has a constant stride.
4118 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4119 const Loop *Lp, ValueToValueMap &StridesMap) {
4120 const Type *Ty = Ptr->getType();
4121 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4123 // Make sure that the pointer does not point to aggregate types.
4124 const PointerType *PtrTy = cast<PointerType>(Ty);
4125 if (PtrTy->getElementType()->isAggregateType()) {
4126 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4131 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4133 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4135 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4136 << *Ptr << " SCEV: " << *PtrScev << "\n");
4140 // The accesss function must stride over the innermost loop.
4141 if (Lp != AR->getLoop()) {
4142 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4143 *Ptr << " SCEV: " << *PtrScev << "\n");
4146 // The address calculation must not wrap. Otherwise, a dependence could be
4148 // An inbounds getelementptr that is a AddRec with a unit stride
4149 // cannot wrap per definition. The unit stride requirement is checked later.
4150 // An getelementptr without an inbounds attribute and unit stride would have
4151 // to access the pointer value "0" which is undefined behavior in address
4152 // space 0, therefore we can also vectorize this case.
4153 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4154 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4155 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4156 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4157 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4158 << *Ptr << " SCEV: " << *PtrScev << "\n");
4162 // Check the step is constant.
4163 const SCEV *Step = AR->getStepRecurrence(*SE);
4165 // Calculate the pointer stride and check if it is consecutive.
4166 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4168 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4169 " SCEV: " << *PtrScev << "\n");
4173 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4174 const APInt &APStepVal = C->getValue()->getValue();
4176 // Huge step value - give up.
4177 if (APStepVal.getBitWidth() > 64)
4180 int64_t StepVal = APStepVal.getSExtValue();
4183 int64_t Stride = StepVal / Size;
4184 int64_t Rem = StepVal % Size;
4188 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4189 // know we can't "wrap around the address space". In case of address space
4190 // zero we know that this won't happen without triggering undefined behavior.
4191 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4192 Stride != 1 && Stride != -1)
4198 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4199 unsigned TypeByteSize) {
4200 // If loads occur at a distance that is not a multiple of a feasible vector
4201 // factor store-load forwarding does not take place.
4202 // Positive dependences might cause troubles because vectorizing them might
4203 // prevent store-load forwarding making vectorized code run a lot slower.
4204 // a[i] = a[i-3] ^ a[i-8];
4205 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4206 // hence on your typical architecture store-load forwarding does not take
4207 // place. Vectorizing in such cases does not make sense.
4208 // Store-load forwarding distance.
4209 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4210 // Maximum vector factor.
4211 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4212 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4213 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4215 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4217 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4218 MaxVFWithoutSLForwardIssues = (vf >>=1);
4223 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4224 DEBUG(dbgs() << "LV: Distance " << Distance <<
4225 " that could cause a store-load forwarding conflict\n");
4229 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4230 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4231 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4235 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4236 const MemAccessInfo &B, unsigned BIdx,
4237 ValueToValueMap &Strides) {
4238 assert (AIdx < BIdx && "Must pass arguments in program order");
4240 Value *APtr = A.getPointer();
4241 Value *BPtr = B.getPointer();
4242 bool AIsWrite = A.getInt();
4243 bool BIsWrite = B.getInt();
4245 // Two reads are independent.
4246 if (!AIsWrite && !BIsWrite)
4249 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4250 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4252 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4253 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4255 const SCEV *Src = AScev;
4256 const SCEV *Sink = BScev;
4258 // If the induction step is negative we have to invert source and sink of the
4260 if (StrideAPtr < 0) {
4263 std::swap(APtr, BPtr);
4264 std::swap(Src, Sink);
4265 std::swap(AIsWrite, BIsWrite);
4266 std::swap(AIdx, BIdx);
4267 std::swap(StrideAPtr, StrideBPtr);
4270 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4272 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4273 << "(Induction step: " << StrideAPtr << ")\n");
4274 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4275 << *InstMap[BIdx] << ": " << *Dist << "\n");
4277 // Need consecutive accesses. We don't want to vectorize
4278 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4279 // the address space.
4280 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4281 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4285 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4287 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4288 ShouldRetryWithRuntimeCheck = true;
4292 Type *ATy = APtr->getType()->getPointerElementType();
4293 Type *BTy = BPtr->getType()->getPointerElementType();
4294 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4296 // Negative distances are not plausible dependencies.
4297 const APInt &Val = C->getValue()->getValue();
4298 if (Val.isNegative()) {
4299 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4300 if (IsTrueDataDependence &&
4301 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4305 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4309 // Write to the same location with the same size.
4310 // Could be improved to assert type sizes are the same (i32 == float, etc).
4314 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4318 assert(Val.isStrictlyPositive() && "Expect a positive value");
4320 // Positive distance bigger than max vectorization factor.
4323 "LV: ReadWrite-Write positive dependency with different types\n");
4327 unsigned Distance = (unsigned) Val.getZExtValue();
4329 // Bail out early if passed-in parameters make vectorization not feasible.
4330 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4331 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4333 // The distance must be bigger than the size needed for a vectorized version
4334 // of the operation and the size of the vectorized operation must not be
4335 // bigger than the currrent maximum size.
4336 if (Distance < 2*TypeByteSize ||
4337 2*TypeByteSize > MaxSafeDepDistBytes ||
4338 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4339 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4340 << Val.getSExtValue() << '\n');
4344 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4345 Distance : MaxSafeDepDistBytes;
4347 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4348 if (IsTrueDataDependence &&
4349 couldPreventStoreLoadForward(Distance, TypeByteSize))
4352 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4353 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4358 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4359 MemAccessInfoSet &CheckDeps,
4360 ValueToValueMap &Strides) {
4362 MaxSafeDepDistBytes = -1U;
4363 while (!CheckDeps.empty()) {
4364 MemAccessInfo CurAccess = *CheckDeps.begin();
4366 // Get the relevant memory access set.
4367 EquivalenceClasses<MemAccessInfo>::iterator I =
4368 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4370 // Check accesses within this set.
4371 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4372 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4374 // Check every access pair.
4376 CheckDeps.erase(*AI);
4377 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4379 // Check every accessing instruction pair in program order.
4380 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4381 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4382 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4383 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4384 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4386 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4397 bool LoopVectorizationLegality::canVectorizeMemory() {
4399 typedef SmallVector<Value*, 16> ValueVector;
4400 typedef SmallPtrSet<Value*, 16> ValueSet;
4402 // Holds the Load and Store *instructions*.
4406 // Holds all the different accesses in the loop.
4407 unsigned NumReads = 0;
4408 unsigned NumReadWrites = 0;
4410 PtrRtCheck.Pointers.clear();
4411 PtrRtCheck.Need = false;
4413 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4414 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4417 for (Loop::block_iterator bb = TheLoop->block_begin(),
4418 be = TheLoop->block_end(); bb != be; ++bb) {
4420 // Scan the BB and collect legal loads and stores.
4421 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4424 // If this is a load, save it. If this instruction can read from memory
4425 // but is not a load, then we quit. Notice that we don't handle function
4426 // calls that read or write.
4427 if (it->mayReadFromMemory()) {
4428 // Many math library functions read the rounding mode. We will only
4429 // vectorize a loop if it contains known function calls that don't set
4430 // the flag. Therefore, it is safe to ignore this read from memory.
4431 CallInst *Call = dyn_cast<CallInst>(it);
4432 if (Call && getIntrinsicIDForCall(Call, TLI))
4435 LoadInst *Ld = dyn_cast<LoadInst>(it);
4436 if (!Ld) return false;
4437 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4438 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4442 Loads.push_back(Ld);
4443 DepChecker.addAccess(Ld);
4447 // Save 'store' instructions. Abort if other instructions write to memory.
4448 if (it->mayWriteToMemory()) {
4449 StoreInst *St = dyn_cast<StoreInst>(it);
4450 if (!St) return false;
4451 if (!St->isSimple() && !IsAnnotatedParallel) {
4452 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4456 Stores.push_back(St);
4457 DepChecker.addAccess(St);
4462 // Now we have two lists that hold the loads and the stores.
4463 // Next, we find the pointers that they use.
4465 // Check if we see any stores. If there are no stores, then we don't
4466 // care if the pointers are *restrict*.
4467 if (!Stores.size()) {
4468 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4472 AccessAnalysis::DepCandidates DependentAccesses;
4473 AccessAnalysis Accesses(DL, DependentAccesses);
4475 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4476 // multiple times on the same object. If the ptr is accessed twice, once
4477 // for read and once for write, it will only appear once (on the write
4478 // list). This is okay, since we are going to check for conflicts between
4479 // writes and between reads and writes, but not between reads and reads.
4482 ValueVector::iterator I, IE;
4483 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4484 StoreInst *ST = cast<StoreInst>(*I);
4485 Value* Ptr = ST->getPointerOperand();
4487 if (isUniform(Ptr)) {
4488 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4492 // If we did *not* see this pointer before, insert it to the read-write
4493 // list. At this phase it is only a 'write' list.
4494 if (Seen.insert(Ptr)) {
4496 Accesses.addStore(Ptr);
4500 if (IsAnnotatedParallel) {
4502 << "LV: A loop annotated parallel, ignore memory dependency "
4507 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4508 LoadInst *LD = cast<LoadInst>(*I);
4509 Value* Ptr = LD->getPointerOperand();
4510 // If we did *not* see this pointer before, insert it to the
4511 // read list. If we *did* see it before, then it is already in
4512 // the read-write list. This allows us to vectorize expressions
4513 // such as A[i] += x; Because the address of A[i] is a read-write
4514 // pointer. This only works if the index of A[i] is consecutive.
4515 // If the address of i is unknown (for example A[B[i]]) then we may
4516 // read a few words, modify, and write a few words, and some of the
4517 // words may be written to the same address.
4518 bool IsReadOnlyPtr = false;
4519 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4521 IsReadOnlyPtr = true;
4523 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4526 // If we write (or read-write) to a single destination and there are no
4527 // other reads in this loop then is it safe to vectorize.
4528 if (NumReadWrites == 1 && NumReads == 0) {
4529 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4533 // Build dependence sets and check whether we need a runtime pointer bounds
4535 Accesses.buildDependenceSets();
4536 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4538 // Find pointers with computable bounds. We are going to use this information
4539 // to place a runtime bound check.
4540 unsigned NumComparisons = 0;
4541 bool CanDoRT = false;
4543 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4546 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4547 " pointer comparisons.\n");
4549 // If we only have one set of dependences to check pointers among we don't
4550 // need a runtime check.
4551 if (NumComparisons == 0 && NeedRTCheck)
4552 NeedRTCheck = false;
4554 // Check that we did not collect too many pointers or found an unsizeable
4556 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4562 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4565 if (NeedRTCheck && !CanDoRT) {
4566 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4567 "the array bounds.\n");
4572 PtrRtCheck.Need = NeedRTCheck;
4574 bool CanVecMem = true;
4575 if (Accesses.isDependencyCheckNeeded()) {
4576 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4577 CanVecMem = DepChecker.areDepsSafe(
4578 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4579 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4581 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4582 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4585 // Clear the dependency checks. We assume they are not needed.
4586 Accesses.resetDepChecks();
4589 PtrRtCheck.Need = true;
4591 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4592 TheLoop, Strides, true);
4593 // Check that we did not collect too many pointers or found an unsizeable
4595 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4596 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4605 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4606 " need a runtime memory check.\n");
4611 static bool hasMultipleUsesOf(Instruction *I,
4612 SmallPtrSet<Instruction *, 8> &Insts) {
4613 unsigned NumUses = 0;
4614 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4615 if (Insts.count(dyn_cast<Instruction>(*Use)))
4624 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4625 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4626 if (!Set.count(dyn_cast<Instruction>(*Use)))
4631 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4632 ReductionKind Kind) {
4633 if (Phi->getNumIncomingValues() != 2)
4636 // Reduction variables are only found in the loop header block.
4637 if (Phi->getParent() != TheLoop->getHeader())
4640 // Obtain the reduction start value from the value that comes from the loop
4642 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4644 // ExitInstruction is the single value which is used outside the loop.
4645 // We only allow for a single reduction value to be used outside the loop.
4646 // This includes users of the reduction, variables (which form a cycle
4647 // which ends in the phi node).
4648 Instruction *ExitInstruction = 0;
4649 // Indicates that we found a reduction operation in our scan.
4650 bool FoundReduxOp = false;
4652 // We start with the PHI node and scan for all of the users of this
4653 // instruction. All users must be instructions that can be used as reduction
4654 // variables (such as ADD). We must have a single out-of-block user. The cycle
4655 // must include the original PHI.
4656 bool FoundStartPHI = false;
4658 // To recognize min/max patterns formed by a icmp select sequence, we store
4659 // the number of instruction we saw from the recognized min/max pattern,
4660 // to make sure we only see exactly the two instructions.
4661 unsigned NumCmpSelectPatternInst = 0;
4662 ReductionInstDesc ReduxDesc(false, 0);
4664 SmallPtrSet<Instruction *, 8> VisitedInsts;
4665 SmallVector<Instruction *, 8> Worklist;
4666 Worklist.push_back(Phi);
4667 VisitedInsts.insert(Phi);
4669 // A value in the reduction can be used:
4670 // - By the reduction:
4671 // - Reduction operation:
4672 // - One use of reduction value (safe).
4673 // - Multiple use of reduction value (not safe).
4675 // - All uses of the PHI must be the reduction (safe).
4676 // - Otherwise, not safe.
4677 // - By one instruction outside of the loop (safe).
4678 // - By further instructions outside of the loop (not safe).
4679 // - By an instruction that is not part of the reduction (not safe).
4681 // * An instruction type other than PHI or the reduction operation.
4682 // * A PHI in the header other than the initial PHI.
4683 while (!Worklist.empty()) {
4684 Instruction *Cur = Worklist.back();
4685 Worklist.pop_back();
4688 // If the instruction has no users then this is a broken chain and can't be
4689 // a reduction variable.
4690 if (Cur->use_empty())
4693 bool IsAPhi = isa<PHINode>(Cur);
4695 // A header PHI use other than the original PHI.
4696 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4699 // Reductions of instructions such as Div, and Sub is only possible if the
4700 // LHS is the reduction variable.
4701 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4702 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4703 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4706 // Any reduction instruction must be of one of the allowed kinds.
4707 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4708 if (!ReduxDesc.IsReduction)
4711 // A reduction operation must only have one use of the reduction value.
4712 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4713 hasMultipleUsesOf(Cur, VisitedInsts))
4716 // All inputs to a PHI node must be a reduction value.
4717 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4720 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4721 isa<SelectInst>(Cur)))
4722 ++NumCmpSelectPatternInst;
4723 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4724 isa<SelectInst>(Cur)))
4725 ++NumCmpSelectPatternInst;
4727 // Check whether we found a reduction operator.
4728 FoundReduxOp |= !IsAPhi;
4730 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4731 // onto the stack. This way we are going to have seen all inputs to PHI
4732 // nodes once we get to them.
4733 SmallVector<Instruction *, 8> NonPHIs;
4734 SmallVector<Instruction *, 8> PHIs;
4735 for (User *U : Cur->users()) {
4736 Instruction *UI = cast<Instruction>(U);
4738 // Check if we found the exit user.
4739 BasicBlock *Parent = UI->getParent();
4740 if (!TheLoop->contains(Parent)) {
4741 // Exit if you find multiple outside users or if the header phi node is
4742 // being used. In this case the user uses the value of the previous
4743 // iteration, in which case we would loose "VF-1" iterations of the
4744 // reduction operation if we vectorize.
4745 if (ExitInstruction != 0 || Cur == Phi)
4748 // The instruction used by an outside user must be the last instruction
4749 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4750 // operations on the value.
4751 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4754 ExitInstruction = Cur;
4758 // Process instructions only once (termination). Each reduction cycle
4759 // value must only be used once, except by phi nodes and min/max
4760 // reductions which are represented as a cmp followed by a select.
4761 ReductionInstDesc IgnoredVal(false, 0);
4762 if (VisitedInsts.insert(UI)) {
4763 if (isa<PHINode>(UI))
4766 NonPHIs.push_back(UI);
4767 } else if (!isa<PHINode>(UI) &&
4768 ((!isa<FCmpInst>(UI) &&
4769 !isa<ICmpInst>(UI) &&
4770 !isa<SelectInst>(UI)) ||
4771 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4774 // Remember that we completed the cycle.
4776 FoundStartPHI = true;
4778 Worklist.append(PHIs.begin(), PHIs.end());
4779 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4782 // This means we have seen one but not the other instruction of the
4783 // pattern or more than just a select and cmp.
4784 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4785 NumCmpSelectPatternInst != 2)
4788 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4791 // We found a reduction var if we have reached the original phi node and we
4792 // only have a single instruction with out-of-loop users.
4794 // This instruction is allowed to have out-of-loop users.
4795 AllowedExit.insert(ExitInstruction);
4797 // Save the description of this reduction variable.
4798 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4799 ReduxDesc.MinMaxKind);
4800 Reductions[Phi] = RD;
4801 // We've ended the cycle. This is a reduction variable if we have an
4802 // outside user and it has a binary op.
4807 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4808 /// pattern corresponding to a min(X, Y) or max(X, Y).
4809 LoopVectorizationLegality::ReductionInstDesc
4810 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4811 ReductionInstDesc &Prev) {
4813 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4814 "Expect a select instruction");
4815 Instruction *Cmp = 0;
4816 SelectInst *Select = 0;
4818 // We must handle the select(cmp()) as a single instruction. Advance to the
4820 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4821 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4822 return ReductionInstDesc(false, I);
4823 return ReductionInstDesc(Select, Prev.MinMaxKind);
4826 // Only handle single use cases for now.
4827 if (!(Select = dyn_cast<SelectInst>(I)))
4828 return ReductionInstDesc(false, I);
4829 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4830 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4831 return ReductionInstDesc(false, I);
4832 if (!Cmp->hasOneUse())
4833 return ReductionInstDesc(false, I);
4838 // Look for a min/max pattern.
4839 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4840 return ReductionInstDesc(Select, MRK_UIntMin);
4841 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4842 return ReductionInstDesc(Select, MRK_UIntMax);
4843 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4844 return ReductionInstDesc(Select, MRK_SIntMax);
4845 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4846 return ReductionInstDesc(Select, MRK_SIntMin);
4847 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4848 return ReductionInstDesc(Select, MRK_FloatMin);
4849 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4850 return ReductionInstDesc(Select, MRK_FloatMax);
4851 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4852 return ReductionInstDesc(Select, MRK_FloatMin);
4853 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4854 return ReductionInstDesc(Select, MRK_FloatMax);
4856 return ReductionInstDesc(false, I);
4859 LoopVectorizationLegality::ReductionInstDesc
4860 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4862 ReductionInstDesc &Prev) {
4863 bool FP = I->getType()->isFloatingPointTy();
4864 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4865 switch (I->getOpcode()) {
4867 return ReductionInstDesc(false, I);
4868 case Instruction::PHI:
4869 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4870 Kind != RK_FloatMinMax))
4871 return ReductionInstDesc(false, I);
4872 return ReductionInstDesc(I, Prev.MinMaxKind);
4873 case Instruction::Sub:
4874 case Instruction::Add:
4875 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4876 case Instruction::Mul:
4877 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4878 case Instruction::And:
4879 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4880 case Instruction::Or:
4881 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4882 case Instruction::Xor:
4883 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4884 case Instruction::FMul:
4885 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4886 case Instruction::FAdd:
4887 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4888 case Instruction::FCmp:
4889 case Instruction::ICmp:
4890 case Instruction::Select:
4891 if (Kind != RK_IntegerMinMax &&
4892 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4893 return ReductionInstDesc(false, I);
4894 return isMinMaxSelectCmpPattern(I, Prev);
4898 LoopVectorizationLegality::InductionKind
4899 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4900 Type *PhiTy = Phi->getType();
4901 // We only handle integer and pointer inductions variables.
4902 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4903 return IK_NoInduction;
4905 // Check that the PHI is consecutive.
4906 const SCEV *PhiScev = SE->getSCEV(Phi);
4907 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4909 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4910 return IK_NoInduction;
4912 const SCEV *Step = AR->getStepRecurrence(*SE);
4914 // Integer inductions need to have a stride of one.
4915 if (PhiTy->isIntegerTy()) {
4917 return IK_IntInduction;
4918 if (Step->isAllOnesValue())
4919 return IK_ReverseIntInduction;
4920 return IK_NoInduction;
4923 // Calculate the pointer stride and check if it is consecutive.
4924 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4926 return IK_NoInduction;
4928 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4929 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4930 if (C->getValue()->equalsInt(Size))
4931 return IK_PtrInduction;
4932 else if (C->getValue()->equalsInt(0 - Size))
4933 return IK_ReversePtrInduction;
4935 return IK_NoInduction;
4938 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4939 Value *In0 = const_cast<Value*>(V);
4940 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4944 return Inductions.count(PN);
4947 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4948 assert(TheLoop->contains(BB) && "Unknown block used");
4950 // Blocks that do not dominate the latch need predication.
4951 BasicBlock* Latch = TheLoop->getLoopLatch();
4952 return !DT->dominates(BB, Latch);
4955 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4956 SmallPtrSet<Value *, 8>& SafePtrs) {
4957 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4958 // We might be able to hoist the load.
4959 if (it->mayReadFromMemory()) {
4960 LoadInst *LI = dyn_cast<LoadInst>(it);
4961 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4965 // We don't predicate stores at the moment.
4966 if (it->mayWriteToMemory()) {
4967 StoreInst *SI = dyn_cast<StoreInst>(it);
4968 // We only support predication of stores in basic blocks with one
4970 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
4971 !SafePtrs.count(SI->getPointerOperand()) ||
4972 !SI->getParent()->getSinglePredecessor())
4978 // Check that we don't have a constant expression that can trap as operand.
4979 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4981 if (Constant *C = dyn_cast<Constant>(*OI))
4986 // The instructions below can trap.
4987 switch (it->getOpcode()) {
4989 case Instruction::UDiv:
4990 case Instruction::SDiv:
4991 case Instruction::URem:
4992 case Instruction::SRem:
5000 LoopVectorizationCostModel::VectorizationFactor
5001 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5003 // Width 1 means no vectorize
5004 VectorizationFactor Factor = { 1U, 0U };
5005 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5006 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5010 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5011 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5015 // Find the trip count.
5016 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5017 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5019 unsigned WidestType = getWidestType();
5020 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5021 unsigned MaxSafeDepDist = -1U;
5022 if (Legal->getMaxSafeDepDistBytes() != -1U)
5023 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5024 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5025 WidestRegister : MaxSafeDepDist);
5026 unsigned MaxVectorSize = WidestRegister / WidestType;
5027 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5028 DEBUG(dbgs() << "LV: The Widest register is: "
5029 << WidestRegister << " bits.\n");
5031 if (MaxVectorSize == 0) {
5032 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5036 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5037 " into one vector!");
5039 unsigned VF = MaxVectorSize;
5041 // If we optimize the program for size, avoid creating the tail loop.
5043 // If we are unable to calculate the trip count then don't try to vectorize.
5045 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5049 // Find the maximum SIMD width that can fit within the trip count.
5050 VF = TC % MaxVectorSize;
5055 // If the trip count that we found modulo the vectorization factor is not
5056 // zero then we require a tail.
5058 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5064 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5065 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5067 Factor.Width = UserVF;
5071 float Cost = expectedCost(1);
5073 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
5074 for (unsigned i=2; i <= VF; i*=2) {
5075 // Notice that the vector loop needs to be executed less times, so
5076 // we need to divide the cost of the vector loops by the width of
5077 // the vector elements.
5078 float VectorCost = expectedCost(i) / (float)i;
5079 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5080 (int)VectorCost << ".\n");
5081 if (VectorCost < Cost) {
5087 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5088 Factor.Width = Width;
5089 Factor.Cost = Width * Cost;
5093 unsigned LoopVectorizationCostModel::getWidestType() {
5094 unsigned MaxWidth = 8;
5097 for (Loop::block_iterator bb = TheLoop->block_begin(),
5098 be = TheLoop->block_end(); bb != be; ++bb) {
5099 BasicBlock *BB = *bb;
5101 // For each instruction in the loop.
5102 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5103 Type *T = it->getType();
5105 // Only examine Loads, Stores and PHINodes.
5106 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5109 // Examine PHI nodes that are reduction variables.
5110 if (PHINode *PN = dyn_cast<PHINode>(it))
5111 if (!Legal->getReductionVars()->count(PN))
5114 // Examine the stored values.
5115 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5116 T = ST->getValueOperand()->getType();
5118 // Ignore loaded pointer types and stored pointer types that are not
5119 // consecutive. However, we do want to take consecutive stores/loads of
5120 // pointer vectors into account.
5121 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5124 MaxWidth = std::max(MaxWidth,
5125 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5133 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5136 unsigned LoopCost) {
5138 // -- The unroll heuristics --
5139 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5140 // There are many micro-architectural considerations that we can't predict
5141 // at this level. For example frontend pressure (on decode or fetch) due to
5142 // code size, or the number and capabilities of the execution ports.
5144 // We use the following heuristics to select the unroll factor:
5145 // 1. If the code has reductions the we unroll in order to break the cross
5146 // iteration dependency.
5147 // 2. If the loop is really small then we unroll in order to reduce the loop
5149 // 3. We don't unroll if we think that we will spill registers to memory due
5150 // to the increased register pressure.
5152 // Use the user preference, unless 'auto' is selected.
5156 // When we optimize for size we don't unroll.
5160 // We used the distance for the unroll factor.
5161 if (Legal->getMaxSafeDepDistBytes() != -1U)
5164 // Do not unroll loops with a relatively small trip count.
5165 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5166 TheLoop->getLoopLatch());
5167 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5170 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5171 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5175 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5176 TargetNumRegisters = ForceTargetNumScalarRegs;
5178 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5179 TargetNumRegisters = ForceTargetNumVectorRegs;
5182 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5183 // We divide by these constants so assume that we have at least one
5184 // instruction that uses at least one register.
5185 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5186 R.NumInstructions = std::max(R.NumInstructions, 1U);
5188 // We calculate the unroll factor using the following formula.
5189 // Subtract the number of loop invariants from the number of available
5190 // registers. These registers are used by all of the unrolled instances.
5191 // Next, divide the remaining registers by the number of registers that is
5192 // required by the loop, in order to estimate how many parallel instances
5193 // fit without causing spills. All of this is rounded down if necessary to be
5194 // a power of two. We want power of two unroll factors to simplify any
5195 // addressing operations or alignment considerations.
5196 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5199 // Don't count the induction variable as unrolled.
5200 if (EnableIndVarRegisterHeur)
5201 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5202 std::max(1U, (R.MaxLocalUsers - 1)));
5204 // Clamp the unroll factor ranges to reasonable factors.
5205 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5207 // Check if the user has overridden the unroll max.
5209 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5210 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5212 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5213 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5216 // If we did not calculate the cost for VF (because the user selected the VF)
5217 // then we calculate the cost of VF here.
5219 LoopCost = expectedCost(VF);
5221 // Clamp the calculated UF to be between the 1 and the max unroll factor
5222 // that the target allows.
5223 if (UF > MaxUnrollSize)
5228 // Unroll if we vectorized this loop and there is a reduction that could
5229 // benefit from unrolling.
5230 if (VF > 1 && Legal->getReductionVars()->size()) {
5231 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5235 // Note that if we've already vectorized the loop we will have done the
5236 // runtime check and so unrolling won't require further checks.
5237 bool UnrollingRequiresRuntimePointerCheck =
5238 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5240 // We want to unroll small loops in order to reduce the loop overhead and
5241 // potentially expose ILP opportunities.
5242 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5243 if (!UnrollingRequiresRuntimePointerCheck &&
5244 LoopCost < SmallLoopCost) {
5245 // We assume that the cost overhead is 1 and we use the cost model
5246 // to estimate the cost of the loop and unroll until the cost of the
5247 // loop overhead is about 5% of the cost of the loop.
5248 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5250 // Unroll until store/load ports (estimated by max unroll factor) are
5252 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5253 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5255 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5256 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5257 return std::max(StoresUF, LoadsUF);
5260 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5264 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5268 LoopVectorizationCostModel::RegisterUsage
5269 LoopVectorizationCostModel::calculateRegisterUsage() {
5270 // This function calculates the register usage by measuring the highest number
5271 // of values that are alive at a single location. Obviously, this is a very
5272 // rough estimation. We scan the loop in a topological order in order and
5273 // assign a number to each instruction. We use RPO to ensure that defs are
5274 // met before their users. We assume that each instruction that has in-loop
5275 // users starts an interval. We record every time that an in-loop value is
5276 // used, so we have a list of the first and last occurrences of each
5277 // instruction. Next, we transpose this data structure into a multi map that
5278 // holds the list of intervals that *end* at a specific location. This multi
5279 // map allows us to perform a linear search. We scan the instructions linearly
5280 // and record each time that a new interval starts, by placing it in a set.
5281 // If we find this value in the multi-map then we remove it from the set.
5282 // The max register usage is the maximum size of the set.
5283 // We also search for instructions that are defined outside the loop, but are
5284 // used inside the loop. We need this number separately from the max-interval
5285 // usage number because when we unroll, loop-invariant values do not take
5287 LoopBlocksDFS DFS(TheLoop);
5291 R.NumInstructions = 0;
5293 // Each 'key' in the map opens a new interval. The values
5294 // of the map are the index of the 'last seen' usage of the
5295 // instruction that is the key.
5296 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5297 // Maps instruction to its index.
5298 DenseMap<unsigned, Instruction*> IdxToInstr;
5299 // Marks the end of each interval.
5300 IntervalMap EndPoint;
5301 // Saves the list of instruction indices that are used in the loop.
5302 SmallSet<Instruction*, 8> Ends;
5303 // Saves the list of values that are used in the loop but are
5304 // defined outside the loop, such as arguments and constants.
5305 SmallPtrSet<Value*, 8> LoopInvariants;
5308 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5309 be = DFS.endRPO(); bb != be; ++bb) {
5310 R.NumInstructions += (*bb)->size();
5311 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5313 Instruction *I = it;
5314 IdxToInstr[Index++] = I;
5316 // Save the end location of each USE.
5317 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5318 Value *U = I->getOperand(i);
5319 Instruction *Instr = dyn_cast<Instruction>(U);
5321 // Ignore non-instruction values such as arguments, constants, etc.
5322 if (!Instr) continue;
5324 // If this instruction is outside the loop then record it and continue.
5325 if (!TheLoop->contains(Instr)) {
5326 LoopInvariants.insert(Instr);
5330 // Overwrite previous end points.
5331 EndPoint[Instr] = Index;
5337 // Saves the list of intervals that end with the index in 'key'.
5338 typedef SmallVector<Instruction*, 2> InstrList;
5339 DenseMap<unsigned, InstrList> TransposeEnds;
5341 // Transpose the EndPoints to a list of values that end at each index.
5342 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5344 TransposeEnds[it->second].push_back(it->first);
5346 SmallSet<Instruction*, 8> OpenIntervals;
5347 unsigned MaxUsage = 0;
5350 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5351 for (unsigned int i = 0; i < Index; ++i) {
5352 Instruction *I = IdxToInstr[i];
5353 // Ignore instructions that are never used within the loop.
5354 if (!Ends.count(I)) continue;
5356 // Remove all of the instructions that end at this location.
5357 InstrList &List = TransposeEnds[i];
5358 for (unsigned int j=0, e = List.size(); j < e; ++j)
5359 OpenIntervals.erase(List[j]);
5361 // Count the number of live interals.
5362 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5364 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5365 OpenIntervals.size() << '\n');
5367 // Add the current instruction to the list of open intervals.
5368 OpenIntervals.insert(I);
5371 unsigned Invariant = LoopInvariants.size();
5372 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5373 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5374 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5376 R.LoopInvariantRegs = Invariant;
5377 R.MaxLocalUsers = MaxUsage;
5381 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5385 for (Loop::block_iterator bb = TheLoop->block_begin(),
5386 be = TheLoop->block_end(); bb != be; ++bb) {
5387 unsigned BlockCost = 0;
5388 BasicBlock *BB = *bb;
5390 // For each instruction in the old loop.
5391 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5392 // Skip dbg intrinsics.
5393 if (isa<DbgInfoIntrinsic>(it))
5396 unsigned C = getInstructionCost(it, VF);
5398 // Check if we should override the cost.
5399 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5400 C = ForceTargetInstructionCost;
5403 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5404 VF << " For instruction: " << *it << '\n');
5407 // We assume that if-converted blocks have a 50% chance of being executed.
5408 // When the code is scalar then some of the blocks are avoided due to CF.
5409 // When the code is vectorized we execute all code paths.
5410 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5419 /// \brief Check whether the address computation for a non-consecutive memory
5420 /// access looks like an unlikely candidate for being merged into the indexing
5423 /// We look for a GEP which has one index that is an induction variable and all
5424 /// other indices are loop invariant. If the stride of this access is also
5425 /// within a small bound we decide that this address computation can likely be
5426 /// merged into the addressing mode.
5427 /// In all other cases, we identify the address computation as complex.
5428 static bool isLikelyComplexAddressComputation(Value *Ptr,
5429 LoopVectorizationLegality *Legal,
5430 ScalarEvolution *SE,
5431 const Loop *TheLoop) {
5432 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5436 // We are looking for a gep with all loop invariant indices except for one
5437 // which should be an induction variable.
5438 unsigned NumOperands = Gep->getNumOperands();
5439 for (unsigned i = 1; i < NumOperands; ++i) {
5440 Value *Opd = Gep->getOperand(i);
5441 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5442 !Legal->isInductionVariable(Opd))
5446 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5447 // can likely be merged into the address computation.
5448 unsigned MaxMergeDistance = 64;
5450 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5454 // Check the step is constant.
5455 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5456 // Calculate the pointer stride and check if it is consecutive.
5457 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5461 const APInt &APStepVal = C->getValue()->getValue();
5463 // Huge step value - give up.
5464 if (APStepVal.getBitWidth() > 64)
5467 int64_t StepVal = APStepVal.getSExtValue();
5469 return StepVal > MaxMergeDistance;
5472 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5473 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5479 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5480 // If we know that this instruction will remain uniform, check the cost of
5481 // the scalar version.
5482 if (Legal->isUniformAfterVectorization(I))
5485 Type *RetTy = I->getType();
5486 Type *VectorTy = ToVectorTy(RetTy, VF);
5488 // TODO: We need to estimate the cost of intrinsic calls.
5489 switch (I->getOpcode()) {
5490 case Instruction::GetElementPtr:
5491 // We mark this instruction as zero-cost because the cost of GEPs in
5492 // vectorized code depends on whether the corresponding memory instruction
5493 // is scalarized or not. Therefore, we handle GEPs with the memory
5494 // instruction cost.
5496 case Instruction::Br: {
5497 return TTI.getCFInstrCost(I->getOpcode());
5499 case Instruction::PHI:
5500 //TODO: IF-converted IFs become selects.
5502 case Instruction::Add:
5503 case Instruction::FAdd:
5504 case Instruction::Sub:
5505 case Instruction::FSub:
5506 case Instruction::Mul:
5507 case Instruction::FMul:
5508 case Instruction::UDiv:
5509 case Instruction::SDiv:
5510 case Instruction::FDiv:
5511 case Instruction::URem:
5512 case Instruction::SRem:
5513 case Instruction::FRem:
5514 case Instruction::Shl:
5515 case Instruction::LShr:
5516 case Instruction::AShr:
5517 case Instruction::And:
5518 case Instruction::Or:
5519 case Instruction::Xor: {
5520 // Since we will replace the stride by 1 the multiplication should go away.
5521 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5523 // Certain instructions can be cheaper to vectorize if they have a constant
5524 // second vector operand. One example of this are shifts on x86.
5525 TargetTransformInfo::OperandValueKind Op1VK =
5526 TargetTransformInfo::OK_AnyValue;
5527 TargetTransformInfo::OperandValueKind Op2VK =
5528 TargetTransformInfo::OK_AnyValue;
5529 Value *Op2 = I->getOperand(1);
5531 // Check for a splat of a constant or for a non uniform vector of constants.
5532 if (isa<ConstantInt>(Op2))
5533 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5534 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5535 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5536 if (cast<Constant>(Op2)->getSplatValue() != NULL)
5537 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5540 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5542 case Instruction::Select: {
5543 SelectInst *SI = cast<SelectInst>(I);
5544 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5545 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5546 Type *CondTy = SI->getCondition()->getType();
5548 CondTy = VectorType::get(CondTy, VF);
5550 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5552 case Instruction::ICmp:
5553 case Instruction::FCmp: {
5554 Type *ValTy = I->getOperand(0)->getType();
5555 VectorTy = ToVectorTy(ValTy, VF);
5556 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5558 case Instruction::Store:
5559 case Instruction::Load: {
5560 StoreInst *SI = dyn_cast<StoreInst>(I);
5561 LoadInst *LI = dyn_cast<LoadInst>(I);
5562 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5564 VectorTy = ToVectorTy(ValTy, VF);
5566 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5567 unsigned AS = SI ? SI->getPointerAddressSpace() :
5568 LI->getPointerAddressSpace();
5569 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5570 // We add the cost of address computation here instead of with the gep
5571 // instruction because only here we know whether the operation is
5574 return TTI.getAddressComputationCost(VectorTy) +
5575 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5577 // Scalarized loads/stores.
5578 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5579 bool Reverse = ConsecutiveStride < 0;
5580 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5581 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5582 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5583 bool IsComplexComputation =
5584 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5586 // The cost of extracting from the value vector and pointer vector.
5587 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5588 for (unsigned i = 0; i < VF; ++i) {
5589 // The cost of extracting the pointer operand.
5590 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5591 // In case of STORE, the cost of ExtractElement from the vector.
5592 // In case of LOAD, the cost of InsertElement into the returned
5594 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5595 Instruction::InsertElement,
5599 // The cost of the scalar loads/stores.
5600 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5601 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5606 // Wide load/stores.
5607 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5608 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5611 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5615 case Instruction::ZExt:
5616 case Instruction::SExt:
5617 case Instruction::FPToUI:
5618 case Instruction::FPToSI:
5619 case Instruction::FPExt:
5620 case Instruction::PtrToInt:
5621 case Instruction::IntToPtr:
5622 case Instruction::SIToFP:
5623 case Instruction::UIToFP:
5624 case Instruction::Trunc:
5625 case Instruction::FPTrunc:
5626 case Instruction::BitCast: {
5627 // We optimize the truncation of induction variable.
5628 // The cost of these is the same as the scalar operation.
5629 if (I->getOpcode() == Instruction::Trunc &&
5630 Legal->isInductionVariable(I->getOperand(0)))
5631 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5632 I->getOperand(0)->getType());
5634 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5635 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5637 case Instruction::Call: {
5638 CallInst *CI = cast<CallInst>(I);
5639 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5640 assert(ID && "Not an intrinsic call!");
5641 Type *RetTy = ToVectorTy(CI->getType(), VF);
5642 SmallVector<Type*, 4> Tys;
5643 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5644 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5645 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5648 // We are scalarizing the instruction. Return the cost of the scalar
5649 // instruction, plus the cost of insert and extract into vector
5650 // elements, times the vector width.
5653 if (!RetTy->isVoidTy() && VF != 1) {
5654 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5656 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5659 // The cost of inserting the results plus extracting each one of the
5661 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5664 // The cost of executing VF copies of the scalar instruction. This opcode
5665 // is unknown. Assume that it is the same as 'mul'.
5666 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5672 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5673 if (Scalar->isVoidTy() || VF == 1)
5675 return VectorType::get(Scalar, VF);
5678 char LoopVectorize::ID = 0;
5679 static const char lv_name[] = "Loop Vectorization";
5680 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5681 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5682 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5683 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5684 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5685 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5686 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5687 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5688 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5691 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5692 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5696 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5697 // Check for a store.
5698 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5699 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5701 // Check for a load.
5702 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5703 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5709 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5710 bool IfPredicateStore) {
5711 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5712 // Holds vector parameters or scalars, in case of uniform vals.
5713 SmallVector<VectorParts, 4> Params;
5715 setDebugLocFromInst(Builder, Instr);
5717 // Find all of the vectorized parameters.
5718 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5719 Value *SrcOp = Instr->getOperand(op);
5721 // If we are accessing the old induction variable, use the new one.
5722 if (SrcOp == OldInduction) {
5723 Params.push_back(getVectorValue(SrcOp));
5727 // Try using previously calculated values.
5728 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5730 // If the src is an instruction that appeared earlier in the basic block
5731 // then it should already be vectorized.
5732 if (SrcInst && OrigLoop->contains(SrcInst)) {
5733 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5734 // The parameter is a vector value from earlier.
5735 Params.push_back(WidenMap.get(SrcInst));
5737 // The parameter is a scalar from outside the loop. Maybe even a constant.
5738 VectorParts Scalars;
5739 Scalars.append(UF, SrcOp);
5740 Params.push_back(Scalars);
5744 assert(Params.size() == Instr->getNumOperands() &&
5745 "Invalid number of operands");
5747 // Does this instruction return a value ?
5748 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5750 Value *UndefVec = IsVoidRetTy ? 0 :
5751 UndefValue::get(Instr->getType());
5752 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5753 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5755 Instruction *InsertPt = Builder.GetInsertPoint();
5756 BasicBlock *IfBlock = Builder.GetInsertBlock();
5757 BasicBlock *CondBlock = 0;
5761 if (IfPredicateStore) {
5762 assert(Instr->getParent()->getSinglePredecessor() &&
5763 "Only support single predecessor blocks");
5764 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5765 Instr->getParent());
5766 VectorLp = LI->getLoopFor(IfBlock);
5767 assert(VectorLp && "Must have a loop for this block");
5770 // For each vector unroll 'part':
5771 for (unsigned Part = 0; Part < UF; ++Part) {
5772 // For each scalar that we create:
5774 // Start an "if (pred) a[i] = ..." block.
5776 if (IfPredicateStore) {
5777 if (Cond[Part]->getType()->isVectorTy())
5779 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5780 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5781 ConstantInt::get(Cond[Part]->getType(), 1));
5782 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5783 LoopVectorBody.push_back(CondBlock);
5784 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5785 // Update Builder with newly created basic block.
5786 Builder.SetInsertPoint(InsertPt);
5789 Instruction *Cloned = Instr->clone();
5791 Cloned->setName(Instr->getName() + ".cloned");
5792 // Replace the operands of the cloned instructions with extracted scalars.
5793 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5794 Value *Op = Params[op][Part];
5795 Cloned->setOperand(op, Op);
5798 // Place the cloned scalar in the new loop.
5799 Builder.Insert(Cloned);
5801 // If the original scalar returns a value we need to place it in a vector
5802 // so that future users will be able to use it.
5804 VecResults[Part] = Cloned;
5807 if (IfPredicateStore) {
5808 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5809 LoopVectorBody.push_back(NewIfBlock);
5810 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5811 Builder.SetInsertPoint(InsertPt);
5812 Instruction *OldBr = IfBlock->getTerminator();
5813 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5814 OldBr->eraseFromParent();
5815 IfBlock = NewIfBlock;
5820 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5821 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5822 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5824 return scalarizeInstruction(Instr, IfPredicateStore);
5827 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5831 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5835 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5837 // When unrolling and the VF is 1, we only need to add a simple scalar.
5838 Type *ITy = Val->getType();
5839 assert(!ITy->isVectorTy() && "Val must be a scalar");
5840 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5841 return Builder.CreateAdd(Val, C, "induction");