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/MapVector.h"
52 #include "llvm/ADT/SetVector.h"
53 #include "llvm/ADT/SmallPtrSet.h"
54 #include "llvm/ADT/SmallSet.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/StringExtras.h"
57 #include "llvm/Analysis/AliasAnalysis.h"
58 #include "llvm/Analysis/Dominators.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/Analysis/Verifier.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/LLVMContext.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/IR/Type.h"
78 #include "llvm/IR/Value.h"
79 #include "llvm/Pass.h"
80 #include "llvm/Support/CommandLine.h"
81 #include "llvm/Support/Debug.h"
82 #include "llvm/Support/PatternMatch.h"
83 #include "llvm/Support/raw_ostream.h"
84 #include "llvm/Support/ValueHandle.h"
85 #include "llvm/Target/TargetLibraryInfo.h"
86 #include "llvm/Transforms/Scalar.h"
87 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
88 #include "llvm/Transforms/Utils/Local.h"
93 using namespace llvm::PatternMatch;
95 static cl::opt<unsigned>
96 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
97 cl::desc("Sets the SIMD width. Zero is autoselect."));
99 static cl::opt<unsigned>
100 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
101 cl::desc("Sets the vectorization unroll count. "
102 "Zero is autoselect."));
105 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
106 cl::desc("Enable if-conversion during vectorization."));
108 /// We don't vectorize loops with a known constant trip count below this number.
109 static cl::opt<unsigned>
110 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
112 cl::desc("Don't vectorize loops with a constant "
113 "trip count that is smaller than this "
116 /// We don't unroll loops with a known constant trip count below this number.
117 static const unsigned TinyTripCountUnrollThreshold = 128;
119 /// When performing memory disambiguation checks at runtime do not make more
120 /// than this number of comparisons.
121 static const unsigned RuntimeMemoryCheckThreshold = 8;
123 /// Maximum simd width.
124 static const unsigned MaxVectorWidth = 64;
126 /// Maximum vectorization unroll count.
127 static const unsigned MaxUnrollFactor = 16;
131 // Forward declarations.
132 class LoopVectorizationLegality;
133 class LoopVectorizationCostModel;
135 /// InnerLoopVectorizer vectorizes loops which contain only one basic
136 /// block to a specified vectorization factor (VF).
137 /// This class performs the widening of scalars into vectors, or multiple
138 /// scalars. This class also implements the following features:
139 /// * It inserts an epilogue loop for handling loops that don't have iteration
140 /// counts that are known to be a multiple of the vectorization factor.
141 /// * It handles the code generation for reduction variables.
142 /// * Scalarization (implementation using scalars) of un-vectorizable
144 /// InnerLoopVectorizer does not perform any vectorization-legality
145 /// checks, and relies on the caller to check for the different legality
146 /// aspects. The InnerLoopVectorizer relies on the
147 /// LoopVectorizationLegality class to provide information about the induction
148 /// and reduction variables that were found to a given vectorization factor.
149 class InnerLoopVectorizer {
151 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
152 DominatorTree *DT, DataLayout *DL,
153 const TargetLibraryInfo *TLI, unsigned VecWidth,
154 unsigned UnrollFactor)
155 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
156 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
157 OldInduction(0), WidenMap(UnrollFactor) {}
159 // Perform the actual loop widening (vectorization).
160 void vectorize(LoopVectorizationLegality *Legal) {
161 // Create a new empty loop. Unlink the old loop and connect the new one.
162 createEmptyLoop(Legal);
163 // Widen each instruction in the old loop to a new one in the new loop.
164 // Use the Legality module to find the induction and reduction variables.
165 vectorizeLoop(Legal);
166 // Register the new loop and update the analysis passes.
171 /// A small list of PHINodes.
172 typedef SmallVector<PHINode*, 4> PhiVector;
173 /// When we unroll loops we have multiple vector values for each scalar.
174 /// This data structure holds the unrolled and vectorized values that
175 /// originated from one scalar instruction.
176 typedef SmallVector<Value*, 2> VectorParts;
178 // When we if-convert we need create edge masks. We have to cache values so
179 // that we don't end up with exponential recursion/IR.
180 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
181 VectorParts> EdgeMaskCache;
183 /// Add code that checks at runtime if the accessed arrays overlap.
184 /// Returns the comparator value or NULL if no check is needed.
185 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
187 /// Create an empty loop, based on the loop ranges of the old loop.
188 void createEmptyLoop(LoopVectorizationLegality *Legal);
189 /// Copy and widen the instructions from the old loop.
190 void vectorizeLoop(LoopVectorizationLegality *Legal);
192 /// A helper function that computes the predicate of the block BB, assuming
193 /// that the header block of the loop is set to True. It returns the *entry*
194 /// mask for the block BB.
195 VectorParts createBlockInMask(BasicBlock *BB);
196 /// A helper function that computes the predicate of the edge between SRC
198 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
200 /// A helper function to vectorize a single BB within the innermost loop.
201 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
204 /// Insert the new loop to the loop hierarchy and pass manager
205 /// and update the analysis passes.
206 void updateAnalysis();
208 /// This instruction is un-vectorizable. Implement it as a sequence
210 void scalarizeInstruction(Instruction *Instr);
212 /// Vectorize Load and Store instructions,
213 void vectorizeMemoryInstruction(Instruction *Instr,
214 LoopVectorizationLegality *Legal);
216 /// Create a broadcast instruction. This method generates a broadcast
217 /// instruction (shuffle) for loop invariant values and for the induction
218 /// value. If this is the induction variable then we extend it to N, N+1, ...
219 /// this is needed because each iteration in the loop corresponds to a SIMD
221 Value *getBroadcastInstrs(Value *V);
223 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
224 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
225 /// The sequence starts at StartIndex.
226 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
228 /// When we go over instructions in the basic block we rely on previous
229 /// values within the current basic block or on loop invariant values.
230 /// When we widen (vectorize) values we place them in the map. If the values
231 /// are not within the map, they have to be loop invariant, so we simply
232 /// broadcast them into a vector.
233 VectorParts &getVectorValue(Value *V);
235 /// Generate a shuffle sequence that will reverse the vector Vec.
236 Value *reverseVector(Value *Vec);
238 /// This is a helper class that holds the vectorizer state. It maps scalar
239 /// instructions to vector instructions. When the code is 'unrolled' then
240 /// then a single scalar value is mapped to multiple vector parts. The parts
241 /// are stored in the VectorPart type.
243 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
245 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
247 /// \return True if 'Key' is saved in the Value Map.
248 bool has(Value *Key) const { return MapStorage.count(Key); }
250 /// Initializes a new entry in the map. Sets all of the vector parts to the
251 /// save value in 'Val'.
252 /// \return A reference to a vector with splat values.
253 VectorParts &splat(Value *Key, Value *Val) {
254 VectorParts &Entry = MapStorage[Key];
255 Entry.assign(UF, Val);
259 ///\return A reference to the value that is stored at 'Key'.
260 VectorParts &get(Value *Key) {
261 VectorParts &Entry = MapStorage[Key];
264 assert(Entry.size() == UF);
269 /// The unroll factor. Each entry in the map stores this number of vector
273 /// Map storage. We use std::map and not DenseMap because insertions to a
274 /// dense map invalidates its iterators.
275 std::map<Value *, VectorParts> MapStorage;
278 /// The original loop.
280 /// Scev analysis to use.
288 /// Target Library Info.
289 const TargetLibraryInfo *TLI;
291 /// The vectorization SIMD factor to use. Each vector will have this many
294 /// The vectorization unroll factor to use. Each scalar is vectorized to this
295 /// many different vector instructions.
298 /// The builder that we use
301 // --- Vectorization state ---
303 /// The vector-loop preheader.
304 BasicBlock *LoopVectorPreHeader;
305 /// The scalar-loop preheader.
306 BasicBlock *LoopScalarPreHeader;
307 /// Middle Block between the vector and the scalar.
308 BasicBlock *LoopMiddleBlock;
309 ///The ExitBlock of the scalar loop.
310 BasicBlock *LoopExitBlock;
311 ///The vector loop body.
312 BasicBlock *LoopVectorBody;
313 ///The scalar loop body.
314 BasicBlock *LoopScalarBody;
315 /// A list of all bypass blocks. The first block is the entry of the loop.
316 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
318 /// The new Induction variable which was added to the new block.
320 /// The induction variable of the old basic block.
321 PHINode *OldInduction;
322 /// Holds the extended (to the widest induction type) start index.
324 /// Maps scalars to widened vectors.
326 EdgeMaskCache MaskCache;
329 /// \brief Set/reset the debug location in the IR builder using the RAII idiom.
330 class DebugLocSetter {
331 IRBuilder<> &Builder;
334 DebugLocSetter(const DebugLocSetter&);
335 DebugLocSetter &operator=(const DebugLocSetter&);
338 /// \brief Set the debug location in the IRBuilder 'B' using the instruction
340 DebugLocSetter(IRBuilder<> &B, Instruction *Inst) : Builder(B) {
341 OldDL = Builder.getCurrentDebugLocation();
342 // Handle null instructions gracefully. This is so we can use a dyn_cast on
343 // values without nowing it is an instruction.
345 Builder.SetCurrentDebugLocation(Inst->getDebugLoc());
349 Builder.SetCurrentDebugLocation(OldDL);
353 /// \brief Look for a meaningful debug location on the instruction or it's
355 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
360 if (I->getDebugLoc() != Empty)
363 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
364 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
365 if (OpInst->getDebugLoc() != Empty)
372 /// \brief Check if conditionally executed loads are hoistable.
374 /// This class has two functions: isHoistableLoad and canHoistAllLoads.
375 /// isHoistableLoad should be called on all load instructions that are executed
376 /// conditionally. After all conditional loads are processed, the client should
377 /// call canHoistAllLoads to determine if all of the conditional executed loads
378 /// have an unconditional memory access to the same memory address in the loop.
380 typedef SmallPtrSet<Value *, 8> MemorySet;
384 MemorySet CondLoadAddrSet;
387 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
389 /// \brief Check if the instruction is a load with a identifiable address.
390 bool isHoistableLoad(Instruction *L);
392 /// \brief Check if all of the conditional loads are hoistable because there
393 /// exists an unconditional memory access to the same address in the loop.
394 bool canHoistAllLoads();
397 bool LoadHoisting::isHoistableLoad(Instruction *L) {
398 LoadInst *LI = dyn_cast<LoadInst>(L);
402 CondLoadAddrSet.insert(LI->getPointerOperand());
406 static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
407 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
408 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
409 Set.insert(LI->getPointerOperand());
410 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
411 Set.insert(SI->getPointerOperand());
415 bool LoadHoisting::canHoistAllLoads() {
416 // No conditional loads.
417 if (CondLoadAddrSet.empty())
420 MemorySet UncondMemAccesses;
421 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
422 BasicBlock *LoopLatch = TheLoop->getLoopLatch();
424 // Iterate over the unconditional blocks and collect memory access addresses.
425 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
426 BasicBlock *BB = LoopBlocks[i];
428 // Ignore conditional blocks.
429 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
432 addMemAccesses(BB, UncondMemAccesses);
435 // And make sure there is a matching unconditional access for every
437 for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
438 ME = CondLoadAddrSet.end(); MI != ME; ++MI)
439 if (!UncondMemAccesses.count(*MI))
445 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
446 /// to what vectorization factor.
447 /// This class does not look at the profitability of vectorization, only the
448 /// legality. This class has two main kinds of checks:
449 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
450 /// will change the order of memory accesses in a way that will change the
451 /// correctness of the program.
452 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
453 /// checks for a number of different conditions, such as the availability of a
454 /// single induction variable, that all types are supported and vectorize-able,
455 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
456 /// This class is also used by InnerLoopVectorizer for identifying
457 /// induction variable and the different reduction variables.
458 class LoopVectorizationLegality {
460 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
461 DominatorTree *DT, TargetLibraryInfo *TLI)
462 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
463 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
464 MaxSafeDepDistBytes(-1U), LoadSpeculation(L, DT) {}
466 /// This enum represents the kinds of reductions that we support.
468 RK_NoReduction, ///< Not a reduction.
469 RK_IntegerAdd, ///< Sum of integers.
470 RK_IntegerMult, ///< Product of integers.
471 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
472 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
473 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
474 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
475 RK_FloatAdd, ///< Sum of floats.
476 RK_FloatMult, ///< Product of floats.
477 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
480 /// This enum represents the kinds of inductions that we support.
482 IK_NoInduction, ///< Not an induction variable.
483 IK_IntInduction, ///< Integer induction variable. Step = 1.
484 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
485 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
486 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
489 // This enum represents the kind of minmax reduction.
490 enum MinMaxReductionKind {
500 /// This POD struct holds information about reduction variables.
501 struct ReductionDescriptor {
502 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
503 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
505 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
506 MinMaxReductionKind MK)
507 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
509 // The starting value of the reduction.
510 // It does not have to be zero!
511 TrackingVH<Value> StartValue;
512 // The instruction who's value is used outside the loop.
513 Instruction *LoopExitInstr;
514 // The kind of the reduction.
516 // If this a min/max reduction the kind of reduction.
517 MinMaxReductionKind MinMaxKind;
520 /// This POD struct holds information about a potential reduction operation.
521 struct ReductionInstDesc {
522 ReductionInstDesc(bool IsRedux, Instruction *I) :
523 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
525 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
526 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
528 // Is this instruction a reduction candidate.
530 // The last instruction in a min/max pattern (select of the select(icmp())
531 // pattern), or the current reduction instruction otherwise.
532 Instruction *PatternLastInst;
533 // If this is a min/max pattern the comparison predicate.
534 MinMaxReductionKind MinMaxKind;
537 // This POD struct holds information about the memory runtime legality
538 // check that a group of pointers do not overlap.
539 struct RuntimePointerCheck {
540 RuntimePointerCheck() : Need(false) {}
542 /// Reset the state of the pointer runtime information.
550 /// Insert a pointer and calculate the start and end SCEVs.
551 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
554 /// This flag indicates if we need to add the runtime check.
556 /// Holds the pointers that we need to check.
557 SmallVector<TrackingVH<Value>, 2> Pointers;
558 /// Holds the pointer value at the beginning of the loop.
559 SmallVector<const SCEV*, 2> Starts;
560 /// Holds the pointer value at the end of the loop.
561 SmallVector<const SCEV*, 2> Ends;
562 /// Holds the information if this pointer is used for writing to memory.
563 SmallVector<bool, 2> IsWritePtr;
564 /// Holds the id of the set of pointers that could be dependent because of a
565 /// shared underlying object.
566 SmallVector<unsigned, 2> DependencySetId;
569 /// A POD for saving information about induction variables.
570 struct InductionInfo {
571 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
572 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
574 TrackingVH<Value> StartValue;
579 /// ReductionList contains the reduction descriptors for all
580 /// of the reductions that were found in the loop.
581 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
583 /// InductionList saves induction variables and maps them to the
584 /// induction descriptor.
585 typedef MapVector<PHINode*, InductionInfo> InductionList;
587 /// Returns true if it is legal to vectorize this loop.
588 /// This does not mean that it is profitable to vectorize this
589 /// loop, only that it is legal to do so.
592 /// Returns the Induction variable.
593 PHINode *getInduction() { return Induction; }
595 /// Returns the reduction variables found in the loop.
596 ReductionList *getReductionVars() { return &Reductions; }
598 /// Returns the induction variables found in the loop.
599 InductionList *getInductionVars() { return &Inductions; }
601 /// Returns the widest induction type.
602 Type *getWidestInductionType() { return WidestIndTy; }
604 /// Returns True if V is an induction variable in this loop.
605 bool isInductionVariable(const Value *V);
607 /// Return true if the block BB needs to be predicated in order for the loop
608 /// to be vectorized.
609 bool blockNeedsPredication(BasicBlock *BB);
611 /// Check if this pointer is consecutive when vectorizing. This happens
612 /// when the last index of the GEP is the induction variable, or that the
613 /// pointer itself is an induction variable.
614 /// This check allows us to vectorize A[idx] into a wide load/store.
616 /// 0 - Stride is unknown or non consecutive.
617 /// 1 - Address is consecutive.
618 /// -1 - Address is consecutive, and decreasing.
619 int isConsecutivePtr(Value *Ptr);
621 /// Returns true if the value V is uniform within the loop.
622 bool isUniform(Value *V);
624 /// Returns true if this instruction will remain scalar after vectorization.
625 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
627 /// Returns the information that we collected about runtime memory check.
628 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
630 /// This function returns the identity element (or neutral element) for
632 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
634 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
637 /// Check if a single basic block loop is vectorizable.
638 /// At this point we know that this is a loop with a constant trip count
639 /// and we only need to check individual instructions.
640 bool canVectorizeInstrs();
642 /// When we vectorize loops we may change the order in which
643 /// we read and write from memory. This method checks if it is
644 /// legal to vectorize the code, considering only memory constrains.
645 /// Returns true if the loop is vectorizable
646 bool canVectorizeMemory();
648 /// Return true if we can vectorize this loop using the IF-conversion
650 bool canVectorizeWithIfConvert();
652 /// Collect the variables that need to stay uniform after vectorization.
653 void collectLoopUniforms();
655 /// Return true if all of the instructions in the block can be speculatively
657 bool blockCanBePredicated(BasicBlock *BB);
659 /// Returns True, if 'Phi' is the kind of reduction variable for type
660 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
661 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
662 /// Returns a struct describing if the instruction 'I' can be a reduction
663 /// variable of type 'Kind'. If the reduction is a min/max pattern of
664 /// select(icmp()) this function advances the instruction pointer 'I' from the
665 /// compare instruction to the select instruction and stores this pointer in
666 /// 'PatternLastInst' member of the returned struct.
667 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
668 ReductionInstDesc &Desc);
669 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
670 /// pattern corresponding to a min(X, Y) or max(X, Y).
671 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
672 ReductionInstDesc &Prev);
673 /// Returns the induction kind of Phi. This function may return NoInduction
674 /// if the PHI is not an induction variable.
675 InductionKind isInductionVariable(PHINode *Phi);
677 /// The loop that we evaluate.
681 /// DataLayout analysis.
685 /// Target Library Info.
686 TargetLibraryInfo *TLI;
688 // --- vectorization state --- //
690 /// Holds the integer induction variable. This is the counter of the
693 /// Holds the reduction variables.
694 ReductionList Reductions;
695 /// Holds all of the induction variables that we found in the loop.
696 /// Notice that inductions don't need to start at zero and that induction
697 /// variables can be pointers.
698 InductionList Inductions;
699 /// Holds the widest induction type encountered.
702 /// Allowed outside users. This holds the reduction
703 /// vars which can be accessed from outside the loop.
704 SmallPtrSet<Value*, 4> AllowedExit;
705 /// This set holds the variables which are known to be uniform after
707 SmallPtrSet<Instruction*, 4> Uniforms;
708 /// We need to check that all of the pointers in this list are disjoint
710 RuntimePointerCheck PtrRtCheck;
711 /// Can we assume the absence of NaNs.
712 bool HasFunNoNaNAttr;
714 unsigned MaxSafeDepDistBytes;
716 /// Utility to determine whether loads can be speculated.
717 LoadHoisting LoadSpeculation;
720 /// LoopVectorizationCostModel - estimates the expected speedups due to
722 /// In many cases vectorization is not profitable. This can happen because of
723 /// a number of reasons. In this class we mainly attempt to predict the
724 /// expected speedup/slowdowns due to the supported instruction set. We use the
725 /// TargetTransformInfo to query the different backends for the cost of
726 /// different operations.
727 class LoopVectorizationCostModel {
729 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
730 LoopVectorizationLegality *Legal,
731 const TargetTransformInfo &TTI,
732 DataLayout *DL, const TargetLibraryInfo *TLI)
733 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
735 /// Information about vectorization costs
736 struct VectorizationFactor {
737 unsigned Width; // Vector width with best cost
738 unsigned Cost; // Cost of the loop with that width
740 /// \return The most profitable vectorization factor and the cost of that VF.
741 /// This method checks every power of two up to VF. If UserVF is not ZERO
742 /// then this vectorization factor will be selected if vectorization is
744 VectorizationFactor selectVectorizationFactor(bool OptForSize,
747 /// \return The size (in bits) of the widest type in the code that
748 /// needs to be vectorized. We ignore values that remain scalar such as
749 /// 64 bit loop indices.
750 unsigned getWidestType();
752 /// \return The most profitable unroll factor.
753 /// If UserUF is non-zero then this method finds the best unroll-factor
754 /// based on register pressure and other parameters.
755 /// VF and LoopCost are the selected vectorization factor and the cost of the
757 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
760 /// \brief A struct that represents some properties of the register usage
762 struct RegisterUsage {
763 /// Holds the number of loop invariant values that are used in the loop.
764 unsigned LoopInvariantRegs;
765 /// Holds the maximum number of concurrent live intervals in the loop.
766 unsigned MaxLocalUsers;
767 /// Holds the number of instructions in the loop.
768 unsigned NumInstructions;
771 /// \return information about the register usage of the loop.
772 RegisterUsage calculateRegisterUsage();
775 /// Returns the expected execution cost. The unit of the cost does
776 /// not matter because we use the 'cost' units to compare different
777 /// vector widths. The cost that is returned is *not* normalized by
778 /// the factor width.
779 unsigned expectedCost(unsigned VF);
781 /// Returns the execution time cost of an instruction for a given vector
782 /// width. Vector width of one means scalar.
783 unsigned getInstructionCost(Instruction *I, unsigned VF);
785 /// A helper function for converting Scalar types to vector types.
786 /// If the incoming type is void, we return void. If the VF is 1, we return
788 static Type* ToVectorTy(Type *Scalar, unsigned VF);
790 /// Returns whether the instruction is a load or store and will be a emitted
791 /// as a vector operation.
792 bool isConsecutiveLoadOrStore(Instruction *I);
794 /// The loop that we evaluate.
798 /// Loop Info analysis.
800 /// Vectorization legality.
801 LoopVectorizationLegality *Legal;
802 /// Vector target information.
803 const TargetTransformInfo &TTI;
804 /// Target data layout information.
806 /// Target Library Info.
807 const TargetLibraryInfo *TLI;
810 /// Utility class for getting and setting loop vectorizer hints in the form
811 /// of loop metadata.
812 struct LoopVectorizeHints {
813 /// Vectorization width.
815 /// Vectorization unroll factor.
818 LoopVectorizeHints(const Loop *L)
819 : Width(VectorizationFactor)
820 , Unroll(VectorizationUnroll)
821 , LoopID(L->getLoopID()) {
823 // The command line options override any loop metadata except for when
824 // width == 1 which is used to indicate the loop is already vectorized.
825 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
826 Width = VectorizationFactor;
827 if (VectorizationUnroll.getNumOccurrences() > 0)
828 Unroll = VectorizationUnroll;
831 /// Return the loop vectorizer metadata prefix.
832 static StringRef Prefix() { return "llvm.vectorizer."; }
834 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
835 SmallVector<Value*, 2> Vals;
836 Vals.push_back(MDString::get(Context, Name));
837 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
838 return MDNode::get(Context, Vals);
841 /// Mark the loop L as already vectorized by setting the width to 1.
842 void setAlreadyVectorized(Loop *L) {
843 LLVMContext &Context = L->getHeader()->getContext();
847 // Create a new loop id with one more operand for the already_vectorized
848 // hint. If the loop already has a loop id then copy the existing operands.
849 SmallVector<Value*, 4> Vals(1);
851 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
852 Vals.push_back(LoopID->getOperand(i));
854 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
856 MDNode *NewLoopID = MDNode::get(Context, Vals);
857 // Set operand 0 to refer to the loop id itself.
858 NewLoopID->replaceOperandWith(0, NewLoopID);
860 L->setLoopID(NewLoopID);
862 LoopID->replaceAllUsesWith(NewLoopID);
870 /// Find hints specified in the loop metadata.
871 void getHints(const Loop *L) {
875 // First operand should refer to the loop id itself.
876 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
877 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
879 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
880 const MDString *S = 0;
881 SmallVector<Value*, 4> Args;
883 // The expected hint is either a MDString or a MDNode with the first
884 // operand a MDString.
885 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
886 if (!MD || MD->getNumOperands() == 0)
888 S = dyn_cast<MDString>(MD->getOperand(0));
889 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
890 Args.push_back(MD->getOperand(i));
892 S = dyn_cast<MDString>(LoopID->getOperand(i));
893 assert(Args.size() == 0 && "too many arguments for MDString");
899 // Check if the hint starts with the vectorizer prefix.
900 StringRef Hint = S->getString();
901 if (!Hint.startswith(Prefix()))
903 // Remove the prefix.
904 Hint = Hint.substr(Prefix().size(), StringRef::npos);
906 if (Args.size() == 1)
907 getHint(Hint, Args[0]);
911 // Check string hint with one operand.
912 void getHint(StringRef Hint, Value *Arg) {
913 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
915 unsigned Val = C->getZExtValue();
917 if (Hint == "width") {
918 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
919 "Invalid width metadata");
921 } else if (Hint == "unroll") {
922 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
923 "Invalid unroll metadata");
926 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
930 /// The LoopVectorize Pass.
931 struct LoopVectorize : public LoopPass {
932 /// Pass identification, replacement for typeid
935 explicit LoopVectorize() : LoopPass(ID) {
936 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
942 TargetTransformInfo *TTI;
944 TargetLibraryInfo *TLI;
946 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
947 // We only vectorize innermost loops.
951 SE = &getAnalysis<ScalarEvolution>();
952 DL = getAnalysisIfAvailable<DataLayout>();
953 LI = &getAnalysis<LoopInfo>();
954 TTI = &getAnalysis<TargetTransformInfo>();
955 DT = &getAnalysis<DominatorTree>();
956 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
959 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
963 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
964 L->getHeader()->getParent()->getName() << "\"\n");
966 LoopVectorizeHints Hints(L);
968 if (Hints.Width == 1) {
969 DEBUG(dbgs() << "LV: Not vectorizing.\n");
973 // Check if it is legal to vectorize the loop.
974 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
975 if (!LVL.canVectorize()) {
976 DEBUG(dbgs() << "LV: Not vectorizing.\n");
980 // Use the cost model.
981 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
983 // Check the function attributes to find out if this function should be
984 // optimized for size.
985 Function *F = L->getHeader()->getParent();
986 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
987 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
988 unsigned FnIndex = AttributeSet::FunctionIndex;
989 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
990 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
993 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
994 "attribute is used.\n");
998 // Select the optimal vectorization factor.
999 LoopVectorizationCostModel::VectorizationFactor VF;
1000 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
1001 // Select the unroll factor.
1002 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1005 if (VF.Width == 1) {
1006 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1010 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
1011 F->getParent()->getModuleIdentifier()<<"\n");
1012 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
1014 // If we decided that it is *legal* to vectorize the loop then do it.
1015 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1018 // Mark the loop as already vectorized to avoid vectorizing again.
1019 Hints.setAlreadyVectorized(L);
1021 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1025 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1026 LoopPass::getAnalysisUsage(AU);
1027 AU.addRequiredID(LoopSimplifyID);
1028 AU.addRequiredID(LCSSAID);
1029 AU.addRequired<DominatorTree>();
1030 AU.addRequired<LoopInfo>();
1031 AU.addRequired<ScalarEvolution>();
1032 AU.addRequired<TargetTransformInfo>();
1033 AU.addPreserved<LoopInfo>();
1034 AU.addPreserved<DominatorTree>();
1039 } // end anonymous namespace
1041 //===----------------------------------------------------------------------===//
1042 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1043 // LoopVectorizationCostModel.
1044 //===----------------------------------------------------------------------===//
1047 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1048 Loop *Lp, Value *Ptr,
1050 unsigned DepSetId) {
1051 const SCEV *Sc = SE->getSCEV(Ptr);
1052 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1053 assert(AR && "Invalid addrec expression");
1054 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1055 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1056 Pointers.push_back(Ptr);
1057 Starts.push_back(AR->getStart());
1058 Ends.push_back(ScEnd);
1059 IsWritePtr.push_back(WritePtr);
1060 DependencySetId.push_back(DepSetId);
1063 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1064 // Save the current insertion location.
1065 Instruction *Loc = Builder.GetInsertPoint();
1067 // We need to place the broadcast of invariant variables outside the loop.
1068 Instruction *Instr = dyn_cast<Instruction>(V);
1069 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1070 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1072 // Place the code for broadcasting invariant variables in the new preheader.
1074 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1076 // Broadcast the scalar into all locations in the vector.
1077 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1079 // Restore the builder insertion point.
1081 Builder.SetInsertPoint(Loc);
1086 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1088 assert(Val->getType()->isVectorTy() && "Must be a vector");
1089 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1090 "Elem must be an integer");
1091 // Create the types.
1092 Type *ITy = Val->getType()->getScalarType();
1093 VectorType *Ty = cast<VectorType>(Val->getType());
1094 int VLen = Ty->getNumElements();
1095 SmallVector<Constant*, 8> Indices;
1097 // Create a vector of consecutive numbers from zero to VF.
1098 for (int i = 0; i < VLen; ++i) {
1099 int64_t Idx = Negate ? (-i) : i;
1100 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1103 // Add the consecutive indices to the vector value.
1104 Constant *Cv = ConstantVector::get(Indices);
1105 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1106 return Builder.CreateAdd(Val, Cv, "induction");
1109 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1110 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1111 // Make sure that the pointer does not point to structs.
1112 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1115 // If this value is a pointer induction variable we know it is consecutive.
1116 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1117 if (Phi && Inductions.count(Phi)) {
1118 InductionInfo II = Inductions[Phi];
1119 if (IK_PtrInduction == II.IK)
1121 else if (IK_ReversePtrInduction == II.IK)
1125 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1129 unsigned NumOperands = Gep->getNumOperands();
1130 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1132 Value *GpPtr = Gep->getPointerOperand();
1133 // If this GEP value is a consecutive pointer induction variable and all of
1134 // the indices are constant then we know it is consecutive. We can
1135 Phi = dyn_cast<PHINode>(GpPtr);
1136 if (Phi && Inductions.count(Phi)) {
1138 // Make sure that the pointer does not point to structs.
1139 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1140 if (GepPtrType->getElementType()->isAggregateType())
1143 // Make sure that all of the index operands are loop invariant.
1144 for (unsigned i = 1; i < NumOperands; ++i)
1145 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1148 InductionInfo II = Inductions[Phi];
1149 if (IK_PtrInduction == II.IK)
1151 else if (IK_ReversePtrInduction == II.IK)
1155 // Check that all of the gep indices are uniform except for the last.
1156 for (unsigned i = 0; i < NumOperands - 1; ++i)
1157 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1160 // We can emit wide load/stores only if the last index is the induction
1162 const SCEV *Last = SE->getSCEV(LastIndex);
1163 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1164 const SCEV *Step = AR->getStepRecurrence(*SE);
1166 // The memory is consecutive because the last index is consecutive
1167 // and all other indices are loop invariant.
1170 if (Step->isAllOnesValue())
1177 bool LoopVectorizationLegality::isUniform(Value *V) {
1178 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1181 InnerLoopVectorizer::VectorParts&
1182 InnerLoopVectorizer::getVectorValue(Value *V) {
1183 assert(V != Induction && "The new induction variable should not be used.");
1184 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1186 // If we have this scalar in the map, return it.
1187 if (WidenMap.has(V))
1188 return WidenMap.get(V);
1190 // If this scalar is unknown, assume that it is a constant or that it is
1191 // loop invariant. Broadcast V and save the value for future uses.
1192 Value *B = getBroadcastInstrs(V);
1193 return WidenMap.splat(V, B);
1196 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1197 assert(Vec->getType()->isVectorTy() && "Invalid type");
1198 SmallVector<Constant*, 8> ShuffleMask;
1199 for (unsigned i = 0; i < VF; ++i)
1200 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1202 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1203 ConstantVector::get(ShuffleMask),
1208 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1209 LoopVectorizationLegality *Legal) {
1210 // Attempt to issue a wide load.
1211 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1212 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1214 assert((LI || SI) && "Invalid Load/Store instruction");
1216 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1217 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1218 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1219 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1220 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1221 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1222 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1224 if (ScalarAllocatedSize != VectorElementSize)
1225 return scalarizeInstruction(Instr);
1227 // If the pointer is loop invariant or if it is non consecutive,
1228 // scalarize the load.
1229 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1230 bool Reverse = ConsecutiveStride < 0;
1231 bool UniformLoad = LI && Legal->isUniform(Ptr);
1232 if (!ConsecutiveStride || UniformLoad)
1233 return scalarizeInstruction(Instr);
1235 Constant *Zero = Builder.getInt32(0);
1236 VectorParts &Entry = WidenMap.get(Instr);
1238 // Handle consecutive loads/stores.
1239 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1240 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1241 DebugLocSetter SetDL(Builder, Gep);
1242 Value *PtrOperand = Gep->getPointerOperand();
1243 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1244 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1246 // Create the new GEP with the new induction variable.
1247 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1248 Gep2->setOperand(0, FirstBasePtr);
1249 Gep2->setName("gep.indvar.base");
1250 Ptr = Builder.Insert(Gep2);
1252 DebugLocSetter SetDL(Builder, Gep);
1253 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1254 OrigLoop) && "Base ptr must be invariant");
1256 // The last index does not have to be the induction. It can be
1257 // consecutive and be a function of the index. For example A[I+1];
1258 unsigned NumOperands = Gep->getNumOperands();
1259 unsigned LastOperand = NumOperands - 1;
1260 // Create the new GEP with the new induction variable.
1261 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1263 for (unsigned i = 0; i < NumOperands; ++i) {
1264 Value *GepOperand = Gep->getOperand(i);
1265 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1267 // Update last index or loop invariant instruction anchored in loop.
1268 if (i == LastOperand ||
1269 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1270 assert((i == LastOperand ||
1271 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1272 "Must be last index or loop invariant");
1274 VectorParts &GEPParts = getVectorValue(GepOperand);
1275 Value *Index = GEPParts[0];
1276 Index = Builder.CreateExtractElement(Index, Zero);
1277 Gep2->setOperand(i, Index);
1278 Gep2->setName("gep.indvar.idx");
1281 Ptr = Builder.Insert(Gep2);
1283 // Use the induction element ptr.
1284 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1285 DebugLocSetter SetDL(Builder, cast<Instruction>(Ptr));
1286 VectorParts &PtrVal = getVectorValue(Ptr);
1287 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1292 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1293 "We do not allow storing to uniform addresses");
1294 DebugLocSetter SetDL(Builder, SI);
1295 // We don't want to update the value in the map as it might be used in
1296 // another expression. So don't use a reference type for "StoredVal".
1297 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1299 for (unsigned Part = 0; Part < UF; ++Part) {
1300 // Calculate the pointer for the specific unroll-part.
1301 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1304 // If we store to reverse consecutive memory locations then we need
1305 // to reverse the order of elements in the stored value.
1306 StoredVal[Part] = reverseVector(StoredVal[Part]);
1307 // If the address is consecutive but reversed, then the
1308 // wide store needs to start at the last vector element.
1309 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1310 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1313 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1314 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1320 assert(LI && "Must have a load instruction");
1321 DebugLocSetter SetDL(Builder, LI);
1322 for (unsigned Part = 0; Part < UF; ++Part) {
1323 // Calculate the pointer for the specific unroll-part.
1324 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1327 // If the address is consecutive but reversed, then the
1328 // wide store needs to start at the last vector element.
1329 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1330 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1333 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1334 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1335 cast<LoadInst>(LI)->setAlignment(Alignment);
1336 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1340 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1341 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1342 // Holds vector parameters or scalars, in case of uniform vals.
1343 SmallVector<VectorParts, 4> Params;
1345 DebugLocSetter SetDL(Builder, Instr);
1347 // Find all of the vectorized parameters.
1348 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1349 Value *SrcOp = Instr->getOperand(op);
1351 // If we are accessing the old induction variable, use the new one.
1352 if (SrcOp == OldInduction) {
1353 Params.push_back(getVectorValue(SrcOp));
1357 // Try using previously calculated values.
1358 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1360 // If the src is an instruction that appeared earlier in the basic block
1361 // then it should already be vectorized.
1362 if (SrcInst && OrigLoop->contains(SrcInst)) {
1363 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1364 // The parameter is a vector value from earlier.
1365 Params.push_back(WidenMap.get(SrcInst));
1367 // The parameter is a scalar from outside the loop. Maybe even a constant.
1368 VectorParts Scalars;
1369 Scalars.append(UF, SrcOp);
1370 Params.push_back(Scalars);
1374 assert(Params.size() == Instr->getNumOperands() &&
1375 "Invalid number of operands");
1377 // Does this instruction return a value ?
1378 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1380 Value *UndefVec = IsVoidRetTy ? 0 :
1381 UndefValue::get(VectorType::get(Instr->getType(), VF));
1382 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1383 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1385 // For each vector unroll 'part':
1386 for (unsigned Part = 0; Part < UF; ++Part) {
1387 // For each scalar that we create:
1388 for (unsigned Width = 0; Width < VF; ++Width) {
1389 Instruction *Cloned = Instr->clone();
1391 Cloned->setName(Instr->getName() + ".cloned");
1392 // Replace the operands of the cloned instrucions with extracted scalars.
1393 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1394 Value *Op = Params[op][Part];
1395 // Param is a vector. Need to extract the right lane.
1396 if (Op->getType()->isVectorTy())
1397 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1398 Cloned->setOperand(op, Op);
1401 // Place the cloned scalar in the new loop.
1402 Builder.Insert(Cloned);
1404 // If the original scalar returns a value we need to place it in a vector
1405 // so that future users will be able to use it.
1407 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1408 Builder.getInt32(Width));
1414 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1416 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1417 Legal->getRuntimePointerCheck();
1419 if (!PtrRtCheck->Need)
1422 unsigned NumPointers = PtrRtCheck->Pointers.size();
1423 SmallVector<TrackingVH<Value> , 2> Starts;
1424 SmallVector<TrackingVH<Value> , 2> Ends;
1426 SCEVExpander Exp(*SE, "induction");
1428 // Use this type for pointer arithmetic.
1429 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1431 for (unsigned i = 0; i < NumPointers; ++i) {
1432 Value *Ptr = PtrRtCheck->Pointers[i];
1433 const SCEV *Sc = SE->getSCEV(Ptr);
1435 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1436 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1438 Starts.push_back(Ptr);
1439 Ends.push_back(Ptr);
1441 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1443 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1444 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1445 Starts.push_back(Start);
1446 Ends.push_back(End);
1450 IRBuilder<> ChkBuilder(Loc);
1451 // Our instructions might fold to a constant.
1452 Value *MemoryRuntimeCheck = 0;
1453 for (unsigned i = 0; i < NumPointers; ++i) {
1454 for (unsigned j = i+1; j < NumPointers; ++j) {
1455 // No need to check if two readonly pointers intersect.
1456 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1459 // Only need to check pointers between two different dependency sets.
1460 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1463 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1464 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1465 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1466 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1468 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1469 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1470 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1471 if (MemoryRuntimeCheck)
1472 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1474 MemoryRuntimeCheck = IsConflict;
1478 // We have to do this trickery because the IRBuilder might fold the check to a
1479 // constant expression in which case there is no Instruction anchored in a
1481 LLVMContext &Ctx = Loc->getContext();
1482 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1483 ConstantInt::getTrue(Ctx));
1484 ChkBuilder.Insert(Check, "memcheck.conflict");
1489 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1491 In this function we generate a new loop. The new loop will contain
1492 the vectorized instructions while the old loop will continue to run the
1495 [ ] <-- vector loop bypass (may consist of multiple blocks).
1498 | [ ] <-- vector pre header.
1502 | [ ]_| <-- vector loop.
1505 >[ ] <--- middle-block.
1508 | [ ] <--- new preheader.
1512 | [ ]_| <-- old scalar loop to handle remainder.
1515 >[ ] <-- exit block.
1519 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1520 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1521 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1522 assert(ExitBlock && "Must have an exit block");
1524 // Some loops have a single integer induction variable, while other loops
1525 // don't. One example is c++ iterators that often have multiple pointer
1526 // induction variables. In the code below we also support a case where we
1527 // don't have a single induction variable.
1528 OldInduction = Legal->getInduction();
1529 Type *IdxTy = Legal->getWidestInductionType();
1531 // Find the loop boundaries.
1532 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1533 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1535 // Get the total trip count from the count by adding 1.
1536 ExitCount = SE->getAddExpr(ExitCount,
1537 SE->getConstant(ExitCount->getType(), 1));
1539 // Expand the trip count and place the new instructions in the preheader.
1540 // Notice that the pre-header does not change, only the loop body.
1541 SCEVExpander Exp(*SE, "induction");
1543 // Count holds the overall loop count (N).
1544 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1545 BypassBlock->getTerminator());
1547 // The loop index does not have to start at Zero. Find the original start
1548 // value from the induction PHI node. If we don't have an induction variable
1549 // then we know that it starts at zero.
1550 Builder.SetInsertPoint(BypassBlock->getTerminator());
1551 Value *StartIdx = ExtendedIdx = OldInduction ?
1552 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1554 ConstantInt::get(IdxTy, 0);
1556 assert(BypassBlock && "Invalid loop structure");
1557 LoopBypassBlocks.push_back(BypassBlock);
1559 // Split the single block loop into the two loop structure described above.
1560 BasicBlock *VectorPH =
1561 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1562 BasicBlock *VecBody =
1563 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1564 BasicBlock *MiddleBlock =
1565 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1566 BasicBlock *ScalarPH =
1567 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1569 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1571 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1573 // Generate the induction variable.
1574 DebugLocSetter SetDL(Builder, getDebugLocFromInstOrOperands(OldInduction));
1575 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1576 // The loop step is equal to the vectorization factor (num of SIMD elements)
1577 // times the unroll factor (num of SIMD instructions).
1578 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1580 // This is the IR builder that we use to add all of the logic for bypassing
1581 // the new vector loop.
1582 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1583 DebugLocSetter SetDLByPass(BypassBuilder,
1584 getDebugLocFromInstOrOperands(OldInduction));
1586 // We may need to extend the index in case there is a type mismatch.
1587 // We know that the count starts at zero and does not overflow.
1588 if (Count->getType() != IdxTy) {
1589 // The exit count can be of pointer type. Convert it to the correct
1591 if (ExitCount->getType()->isPointerTy())
1592 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1594 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1597 // Add the start index to the loop count to get the new end index.
1598 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1600 // Now we need to generate the expression for N - (N % VF), which is
1601 // the part that the vectorized body will execute.
1602 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1603 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1604 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1605 "end.idx.rnd.down");
1607 // Now, compare the new count to zero. If it is zero skip the vector loop and
1608 // jump to the scalar loop.
1609 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1612 BasicBlock *LastBypassBlock = BypassBlock;
1614 // Generate the code that checks in runtime if arrays overlap. We put the
1615 // checks into a separate block to make the more common case of few elements
1617 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1618 BypassBlock->getTerminator());
1619 if (MemRuntimeCheck) {
1620 // Create a new block containing the memory check.
1621 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1623 LoopBypassBlocks.push_back(CheckBlock);
1625 // Replace the branch into the memory check block with a conditional branch
1626 // for the "few elements case".
1627 Instruction *OldTerm = BypassBlock->getTerminator();
1628 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1629 OldTerm->eraseFromParent();
1631 Cmp = MemRuntimeCheck;
1632 LastBypassBlock = CheckBlock;
1635 LastBypassBlock->getTerminator()->eraseFromParent();
1636 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1639 // We are going to resume the execution of the scalar loop.
1640 // Go over all of the induction variables that we found and fix the
1641 // PHIs that are left in the scalar version of the loop.
1642 // The starting values of PHI nodes depend on the counter of the last
1643 // iteration in the vectorized loop.
1644 // If we come from a bypass edge then we need to start from the original
1647 // This variable saves the new starting index for the scalar loop.
1648 PHINode *ResumeIndex = 0;
1649 LoopVectorizationLegality::InductionList::iterator I, E;
1650 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1651 // Set builder to point to last bypass block.
1652 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1653 for (I = List->begin(), E = List->end(); I != E; ++I) {
1654 PHINode *OrigPhi = I->first;
1655 LoopVectorizationLegality::InductionInfo II = I->second;
1657 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1658 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1659 MiddleBlock->getTerminator());
1660 // We might have extended the type of the induction variable but we need a
1661 // truncated version for the scalar loop.
1662 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1663 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1664 MiddleBlock->getTerminator()) : 0;
1666 Value *EndValue = 0;
1668 case LoopVectorizationLegality::IK_NoInduction:
1669 llvm_unreachable("Unknown induction");
1670 case LoopVectorizationLegality::IK_IntInduction: {
1671 // Handle the integer induction counter.
1672 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1674 // We have the canonical induction variable.
1675 if (OrigPhi == OldInduction) {
1676 // Create a truncated version of the resume value for the scalar loop,
1677 // we might have promoted the type to a larger width.
1679 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1680 // The new PHI merges the original incoming value, in case of a bypass,
1681 // or the value at the end of the vectorized loop.
1682 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1683 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1684 TruncResumeVal->addIncoming(EndValue, VecBody);
1686 // We know what the end value is.
1687 EndValue = IdxEndRoundDown;
1688 // We also know which PHI node holds it.
1689 ResumeIndex = ResumeVal;
1693 // Not the canonical induction variable - add the vector loop count to the
1695 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1696 II.StartValue->getType(),
1698 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1701 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1702 // Convert the CountRoundDown variable to the PHI size.
1703 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1704 II.StartValue->getType(),
1706 // Handle reverse integer induction counter.
1707 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1710 case LoopVectorizationLegality::IK_PtrInduction: {
1711 // For pointer induction variables, calculate the offset using
1713 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1717 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1718 // The value at the end of the loop for the reverse pointer is calculated
1719 // by creating a GEP with a negative index starting from the start value.
1720 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1721 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1723 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1729 // The new PHI merges the original incoming value, in case of a bypass,
1730 // or the value at the end of the vectorized loop.
1731 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1732 if (OrigPhi == OldInduction)
1733 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1735 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1737 ResumeVal->addIncoming(EndValue, VecBody);
1739 // Fix the scalar body counter (PHI node).
1740 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1741 // The old inductions phi node in the scalar body needs the truncated value.
1742 if (OrigPhi == OldInduction)
1743 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1745 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1748 // If we are generating a new induction variable then we also need to
1749 // generate the code that calculates the exit value. This value is not
1750 // simply the end of the counter because we may skip the vectorized body
1751 // in case of a runtime check.
1753 assert(!ResumeIndex && "Unexpected resume value found");
1754 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1755 MiddleBlock->getTerminator());
1756 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1757 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1758 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1761 // Make sure that we found the index where scalar loop needs to continue.
1762 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1763 "Invalid resume Index");
1765 // Add a check in the middle block to see if we have completed
1766 // all of the iterations in the first vector loop.
1767 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1768 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1769 ResumeIndex, "cmp.n",
1770 MiddleBlock->getTerminator());
1772 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1773 // Remove the old terminator.
1774 MiddleBlock->getTerminator()->eraseFromParent();
1776 // Create i+1 and fill the PHINode.
1777 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1778 Induction->addIncoming(StartIdx, VectorPH);
1779 Induction->addIncoming(NextIdx, VecBody);
1780 // Create the compare.
1781 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1782 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1784 // Now we have two terminators. Remove the old one from the block.
1785 VecBody->getTerminator()->eraseFromParent();
1787 // Get ready to start creating new instructions into the vectorized body.
1788 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1790 // Create and register the new vector loop.
1791 Loop* Lp = new Loop();
1792 Loop *ParentLoop = OrigLoop->getParentLoop();
1794 // Insert the new loop into the loop nest and register the new basic blocks.
1796 ParentLoop->addChildLoop(Lp);
1797 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1798 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1799 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1800 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1801 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1803 LI->addTopLevelLoop(Lp);
1806 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1809 LoopVectorPreHeader = VectorPH;
1810 LoopScalarPreHeader = ScalarPH;
1811 LoopMiddleBlock = MiddleBlock;
1812 LoopExitBlock = ExitBlock;
1813 LoopVectorBody = VecBody;
1814 LoopScalarBody = OldBasicBlock;
1817 /// This function returns the identity element (or neutral element) for
1818 /// the operation K.
1820 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1825 // Adding, Xoring, Oring zero to a number does not change it.
1826 return ConstantInt::get(Tp, 0);
1827 case RK_IntegerMult:
1828 // Multiplying a number by 1 does not change it.
1829 return ConstantInt::get(Tp, 1);
1831 // AND-ing a number with an all-1 value does not change it.
1832 return ConstantInt::get(Tp, -1, true);
1834 // Multiplying a number by 1 does not change it.
1835 return ConstantFP::get(Tp, 1.0L);
1837 // Adding zero to a number does not change it.
1838 return ConstantFP::get(Tp, 0.0L);
1840 llvm_unreachable("Unknown reduction kind");
1844 static Intrinsic::ID
1845 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1846 // If we have an intrinsic call, check if it is trivially vectorizable.
1847 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1848 switch (II->getIntrinsicID()) {
1849 case Intrinsic::sqrt:
1850 case Intrinsic::sin:
1851 case Intrinsic::cos:
1852 case Intrinsic::exp:
1853 case Intrinsic::exp2:
1854 case Intrinsic::log:
1855 case Intrinsic::log10:
1856 case Intrinsic::log2:
1857 case Intrinsic::fabs:
1858 case Intrinsic::floor:
1859 case Intrinsic::ceil:
1860 case Intrinsic::trunc:
1861 case Intrinsic::rint:
1862 case Intrinsic::nearbyint:
1863 case Intrinsic::pow:
1864 case Intrinsic::fma:
1865 case Intrinsic::fmuladd:
1866 return II->getIntrinsicID();
1868 return Intrinsic::not_intrinsic;
1873 return Intrinsic::not_intrinsic;
1876 Function *F = CI->getCalledFunction();
1877 // We're going to make assumptions on the semantics of the functions, check
1878 // that the target knows that it's available in this environment.
1879 if (!F || !TLI->getLibFunc(F->getName(), Func))
1880 return Intrinsic::not_intrinsic;
1882 // Otherwise check if we have a call to a function that can be turned into a
1883 // vector intrinsic.
1890 return Intrinsic::sin;
1894 return Intrinsic::cos;
1898 return Intrinsic::exp;
1900 case LibFunc::exp2f:
1901 case LibFunc::exp2l:
1902 return Intrinsic::exp2;
1906 return Intrinsic::log;
1907 case LibFunc::log10:
1908 case LibFunc::log10f:
1909 case LibFunc::log10l:
1910 return Intrinsic::log10;
1912 case LibFunc::log2f:
1913 case LibFunc::log2l:
1914 return Intrinsic::log2;
1916 case LibFunc::fabsf:
1917 case LibFunc::fabsl:
1918 return Intrinsic::fabs;
1919 case LibFunc::floor:
1920 case LibFunc::floorf:
1921 case LibFunc::floorl:
1922 return Intrinsic::floor;
1924 case LibFunc::ceilf:
1925 case LibFunc::ceill:
1926 return Intrinsic::ceil;
1927 case LibFunc::trunc:
1928 case LibFunc::truncf:
1929 case LibFunc::truncl:
1930 return Intrinsic::trunc;
1932 case LibFunc::rintf:
1933 case LibFunc::rintl:
1934 return Intrinsic::rint;
1935 case LibFunc::nearbyint:
1936 case LibFunc::nearbyintf:
1937 case LibFunc::nearbyintl:
1938 return Intrinsic::nearbyint;
1942 return Intrinsic::pow;
1945 return Intrinsic::not_intrinsic;
1948 /// This function translates the reduction kind to an LLVM binary operator.
1950 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1952 case LoopVectorizationLegality::RK_IntegerAdd:
1953 return Instruction::Add;
1954 case LoopVectorizationLegality::RK_IntegerMult:
1955 return Instruction::Mul;
1956 case LoopVectorizationLegality::RK_IntegerOr:
1957 return Instruction::Or;
1958 case LoopVectorizationLegality::RK_IntegerAnd:
1959 return Instruction::And;
1960 case LoopVectorizationLegality::RK_IntegerXor:
1961 return Instruction::Xor;
1962 case LoopVectorizationLegality::RK_FloatMult:
1963 return Instruction::FMul;
1964 case LoopVectorizationLegality::RK_FloatAdd:
1965 return Instruction::FAdd;
1966 case LoopVectorizationLegality::RK_IntegerMinMax:
1967 return Instruction::ICmp;
1968 case LoopVectorizationLegality::RK_FloatMinMax:
1969 return Instruction::FCmp;
1971 llvm_unreachable("Unknown reduction operation");
1975 Value *createMinMaxOp(IRBuilder<> &Builder,
1976 LoopVectorizationLegality::MinMaxReductionKind RK,
1979 CmpInst::Predicate P = CmpInst::ICMP_NE;
1982 llvm_unreachable("Unknown min/max reduction kind");
1983 case LoopVectorizationLegality::MRK_UIntMin:
1984 P = CmpInst::ICMP_ULT;
1986 case LoopVectorizationLegality::MRK_UIntMax:
1987 P = CmpInst::ICMP_UGT;
1989 case LoopVectorizationLegality::MRK_SIntMin:
1990 P = CmpInst::ICMP_SLT;
1992 case LoopVectorizationLegality::MRK_SIntMax:
1993 P = CmpInst::ICMP_SGT;
1995 case LoopVectorizationLegality::MRK_FloatMin:
1996 P = CmpInst::FCMP_OLT;
1998 case LoopVectorizationLegality::MRK_FloatMax:
1999 P = CmpInst::FCMP_OGT;
2004 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
2005 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2007 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2009 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2014 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2015 //===------------------------------------------------===//
2017 // Notice: any optimization or new instruction that go
2018 // into the code below should be also be implemented in
2021 //===------------------------------------------------===//
2022 Constant *Zero = Builder.getInt32(0);
2024 // In order to support reduction variables we need to be able to vectorize
2025 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2026 // stages. First, we create a new vector PHI node with no incoming edges.
2027 // We use this value when we vectorize all of the instructions that use the
2028 // PHI. Next, after all of the instructions in the block are complete we
2029 // add the new incoming edges to the PHI. At this point all of the
2030 // instructions in the basic block are vectorized, so we can use them to
2031 // construct the PHI.
2032 PhiVector RdxPHIsToFix;
2034 // Scan the loop in a topological order to ensure that defs are vectorized
2036 LoopBlocksDFS DFS(OrigLoop);
2039 // Vectorize all of the blocks in the original loop.
2040 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2041 be = DFS.endRPO(); bb != be; ++bb)
2042 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2044 // At this point every instruction in the original loop is widened to
2045 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2046 // that we vectorized. The PHI nodes are currently empty because we did
2047 // not want to introduce cycles. Notice that the remaining PHI nodes
2048 // that we need to fix are reduction variables.
2050 // Create the 'reduced' values for each of the induction vars.
2051 // The reduced values are the vector values that we scalarize and combine
2052 // after the loop is finished.
2053 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2055 PHINode *RdxPhi = *it;
2056 assert(RdxPhi && "Unable to recover vectorized PHI");
2058 // Find the reduction variable descriptor.
2059 assert(Legal->getReductionVars()->count(RdxPhi) &&
2060 "Unable to find the reduction variable");
2061 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2062 (*Legal->getReductionVars())[RdxPhi];
2064 // We need to generate a reduction vector from the incoming scalar.
2065 // To do so, we need to generate the 'identity' vector and overide
2066 // one of the elements with the incoming scalar reduction. We need
2067 // to do it in the vector-loop preheader.
2068 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2070 // This is the vector-clone of the value that leaves the loop.
2071 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2072 Type *VecTy = VectorExit[0]->getType();
2074 // Find the reduction identity variable. Zero for addition, or, xor,
2075 // one for multiplication, -1 for And.
2078 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2079 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2080 // MinMax reduction have the start value as their identify.
2081 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2085 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2086 VecTy->getScalarType());
2087 Identity = ConstantVector::getSplat(VF, Iden);
2089 // This vector is the Identity vector where the first element is the
2090 // incoming scalar reduction.
2091 VectorStart = Builder.CreateInsertElement(Identity,
2092 RdxDesc.StartValue, Zero);
2095 // Fix the vector-loop phi.
2096 // We created the induction variable so we know that the
2097 // preheader is the first entry.
2098 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2100 // Reductions do not have to start at zero. They can start with
2101 // any loop invariant values.
2102 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2103 BasicBlock *Latch = OrigLoop->getLoopLatch();
2104 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2105 VectorParts &Val = getVectorValue(LoopVal);
2106 for (unsigned part = 0; part < UF; ++part) {
2107 // Make sure to add the reduction stat value only to the
2108 // first unroll part.
2109 Value *StartVal = (part == 0) ? VectorStart : Identity;
2110 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2111 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2114 // Before each round, move the insertion point right between
2115 // the PHIs and the values we are going to write.
2116 // This allows us to write both PHINodes and the extractelement
2118 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2120 VectorParts RdxParts;
2121 for (unsigned part = 0; part < UF; ++part) {
2122 // This PHINode contains the vectorized reduction variable, or
2123 // the initial value vector, if we bypass the vector loop.
2124 DebugLocSetter SetDL(Builder, RdxDesc.LoopExitInstr);
2126 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2127 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2128 Value *StartVal = (part == 0) ? VectorStart : Identity;
2129 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2130 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2131 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2132 RdxParts.push_back(NewPhi);
2135 // Reduce all of the unrolled parts into a single vector.
2136 Value *ReducedPartRdx = RdxParts[0];
2137 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2138 for (unsigned part = 1; part < UF; ++part) {
2139 DebugLocSetter SetDL(Builder, dyn_cast<Instruction>(RdxParts[part]));
2141 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2142 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2143 RdxParts[part], ReducedPartRdx,
2146 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2147 ReducedPartRdx, RdxParts[part]);
2150 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2151 // and vector ops, reducing the set of values being computed by half each
2153 assert(isPowerOf2_32(VF) &&
2154 "Reduction emission only supported for pow2 vectors!");
2155 Value *TmpVec = ReducedPartRdx;
2156 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2157 for (unsigned i = VF; i != 1; i >>= 1) {
2158 DebugLocSetter SetDL(Builder, dyn_cast<Instruction>(ReducedPartRdx));
2159 // Move the upper half of the vector to the lower half.
2160 for (unsigned j = 0; j != i/2; ++j)
2161 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2163 // Fill the rest of the mask with undef.
2164 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2165 UndefValue::get(Builder.getInt32Ty()));
2168 Builder.CreateShuffleVector(TmpVec,
2169 UndefValue::get(TmpVec->getType()),
2170 ConstantVector::get(ShuffleMask),
2173 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2174 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2177 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2180 // The result is in the first element of the vector.
2183 DebugLocSetter SetDL(Builder, dyn_cast<Instruction>(ReducedPartRdx));
2184 Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2187 // Now, we need to fix the users of the reduction variable
2188 // inside and outside of the scalar remainder loop.
2189 // We know that the loop is in LCSSA form. We need to update the
2190 // PHI nodes in the exit blocks.
2191 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2192 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2193 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2194 if (!LCSSAPhi) continue;
2196 // All PHINodes need to have a single entry edge, or two if
2197 // we already fixed them.
2198 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2200 // We found our reduction value exit-PHI. Update it with the
2201 // incoming bypass edge.
2202 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2203 // Add an edge coming from the bypass.
2204 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2207 }// end of the LCSSA phi scan.
2209 // Fix the scalar loop reduction variable with the incoming reduction sum
2210 // from the vector body and from the backedge value.
2211 int IncomingEdgeBlockIdx =
2212 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2213 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2214 // Pick the other block.
2215 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2216 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2217 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2218 }// end of for each redux variable.
2220 // The Loop exit block may have single value PHI nodes where the incoming
2221 // value is 'undef'. While vectorizing we only handled real values that
2222 // were defined inside the loop. Here we handle the 'undef case'.
2224 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2225 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2226 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2227 if (!LCSSAPhi) continue;
2228 if (LCSSAPhi->getNumIncomingValues() == 1)
2229 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2234 InnerLoopVectorizer::VectorParts
2235 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2236 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2239 // Look for cached value.
2240 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2241 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2242 if (ECEntryIt != MaskCache.end())
2243 return ECEntryIt->second;
2245 VectorParts SrcMask = createBlockInMask(Src);
2247 // The terminator has to be a branch inst!
2248 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2249 assert(BI && "Unexpected terminator found");
2251 if (BI->isConditional()) {
2252 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2254 if (BI->getSuccessor(0) != Dst)
2255 for (unsigned part = 0; part < UF; ++part)
2256 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2258 for (unsigned part = 0; part < UF; ++part)
2259 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2261 MaskCache[Edge] = EdgeMask;
2265 MaskCache[Edge] = SrcMask;
2269 InnerLoopVectorizer::VectorParts
2270 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2271 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2273 // Loop incoming mask is all-one.
2274 if (OrigLoop->getHeader() == BB) {
2275 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2276 return getVectorValue(C);
2279 // This is the block mask. We OR all incoming edges, and with zero.
2280 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2281 VectorParts BlockMask = getVectorValue(Zero);
2284 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2285 VectorParts EM = createEdgeMask(*it, BB);
2286 for (unsigned part = 0; part < UF; ++part)
2287 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2294 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2295 BasicBlock *BB, PhiVector *PV) {
2296 // For each instruction in the old loop.
2297 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2298 VectorParts &Entry = WidenMap.get(it);
2299 switch (it->getOpcode()) {
2300 case Instruction::Br:
2301 // Nothing to do for PHIs and BR, since we already took care of the
2302 // loop control flow instructions.
2304 case Instruction::PHI:{
2305 PHINode* P = cast<PHINode>(it);
2306 // Handle reduction variables:
2307 if (Legal->getReductionVars()->count(P)) {
2308 for (unsigned part = 0; part < UF; ++part) {
2309 // This is phase one of vectorizing PHIs.
2310 Type *VecTy = VectorType::get(it->getType(), VF);
2311 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2312 LoopVectorBody-> getFirstInsertionPt());
2318 // Check for PHI nodes that are lowered to vector selects.
2319 if (P->getParent() != OrigLoop->getHeader()) {
2320 DebugLocSetter SetDL(Builder, P);
2321 // We know that all PHIs in non header blocks are converted into
2322 // selects, so we don't have to worry about the insertion order and we
2323 // can just use the builder.
2324 // At this point we generate the predication tree. There may be
2325 // duplications since this is a simple recursive scan, but future
2326 // optimizations will clean it up.
2328 unsigned NumIncoming = P->getNumIncomingValues();
2330 // Generate a sequence of selects of the form:
2331 // SELECT(Mask3, In3,
2332 // SELECT(Mask2, In2,
2334 for (unsigned In = 0; In < NumIncoming; In++) {
2335 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2337 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2339 for (unsigned part = 0; part < UF; ++part) {
2340 // We might have single edge PHIs (blocks) - use an identity
2341 // 'select' for the first PHI operand.
2343 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2346 // Select between the current value and the previous incoming edge
2347 // based on the incoming mask.
2348 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2349 Entry[part], "predphi");
2355 // This PHINode must be an induction variable.
2356 // Make sure that we know about it.
2357 assert(Legal->getInductionVars()->count(P) &&
2358 "Not an induction variable");
2360 LoopVectorizationLegality::InductionInfo II =
2361 Legal->getInductionVars()->lookup(P);
2363 DebugLocSetter SetDL(Builder, P);
2366 case LoopVectorizationLegality::IK_NoInduction:
2367 llvm_unreachable("Unknown induction");
2368 case LoopVectorizationLegality::IK_IntInduction: {
2369 assert(P->getType() == II.StartValue->getType() && "Types must match");
2370 Type *PhiTy = P->getType();
2372 if (P == OldInduction) {
2373 // Handle the canonical induction variable. We might have had to
2375 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2377 // Handle other induction variables that are now based on the
2379 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2381 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2382 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2385 Broadcasted = getBroadcastInstrs(Broadcasted);
2386 // After broadcasting the induction variable we need to make the vector
2387 // consecutive by adding 0, 1, 2, etc.
2388 for (unsigned part = 0; part < UF; ++part)
2389 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2392 case LoopVectorizationLegality::IK_ReverseIntInduction:
2393 case LoopVectorizationLegality::IK_PtrInduction:
2394 case LoopVectorizationLegality::IK_ReversePtrInduction:
2395 // Handle reverse integer and pointer inductions.
2396 Value *StartIdx = ExtendedIdx;
2397 // This is the normalized GEP that starts counting at zero.
2398 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2401 // Handle the reverse integer induction variable case.
2402 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2403 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2404 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2406 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2409 // This is a new value so do not hoist it out.
2410 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2411 // After broadcasting the induction variable we need to make the
2412 // vector consecutive by adding ... -3, -2, -1, 0.
2413 for (unsigned part = 0; part < UF; ++part)
2414 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2419 // Handle the pointer induction variable case.
2420 assert(P->getType()->isPointerTy() && "Unexpected type.");
2422 // Is this a reverse induction ptr or a consecutive induction ptr.
2423 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2426 // This is the vector of results. Notice that we don't generate
2427 // vector geps because scalar geps result in better code.
2428 for (unsigned part = 0; part < UF; ++part) {
2429 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2430 for (unsigned int i = 0; i < VF; ++i) {
2431 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2432 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2435 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2437 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2439 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2441 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2442 Builder.getInt32(i),
2445 Entry[part] = VecVal;
2452 case Instruction::Add:
2453 case Instruction::FAdd:
2454 case Instruction::Sub:
2455 case Instruction::FSub:
2456 case Instruction::Mul:
2457 case Instruction::FMul:
2458 case Instruction::UDiv:
2459 case Instruction::SDiv:
2460 case Instruction::FDiv:
2461 case Instruction::URem:
2462 case Instruction::SRem:
2463 case Instruction::FRem:
2464 case Instruction::Shl:
2465 case Instruction::LShr:
2466 case Instruction::AShr:
2467 case Instruction::And:
2468 case Instruction::Or:
2469 case Instruction::Xor: {
2470 // Just widen binops.
2471 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2472 DebugLocSetter SetDL(Builder, BinOp);
2473 VectorParts &A = getVectorValue(it->getOperand(0));
2474 VectorParts &B = getVectorValue(it->getOperand(1));
2476 // Use this vector value for all users of the original instruction.
2477 for (unsigned Part = 0; Part < UF; ++Part) {
2478 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2480 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2481 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2482 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2483 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2484 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2486 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2487 VecOp->setIsExact(BinOp->isExact());
2493 case Instruction::Select: {
2495 // If the selector is loop invariant we can create a select
2496 // instruction with a scalar condition. Otherwise, use vector-select.
2497 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2499 DebugLocSetter SetDL(Builder, it);
2501 // The condition can be loop invariant but still defined inside the
2502 // loop. This means that we can't just use the original 'cond' value.
2503 // We have to take the 'vectorized' value and pick the first lane.
2504 // Instcombine will make this a no-op.
2505 VectorParts &Cond = getVectorValue(it->getOperand(0));
2506 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2507 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2508 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2509 Builder.getInt32(0));
2510 for (unsigned Part = 0; Part < UF; ++Part) {
2511 Entry[Part] = Builder.CreateSelect(
2512 InvariantCond ? ScalarCond : Cond[Part],
2519 case Instruction::ICmp:
2520 case Instruction::FCmp: {
2521 // Widen compares. Generate vector compares.
2522 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2523 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2524 DebugLocSetter SetDL(Builder, it);
2525 VectorParts &A = getVectorValue(it->getOperand(0));
2526 VectorParts &B = getVectorValue(it->getOperand(1));
2527 for (unsigned Part = 0; Part < UF; ++Part) {
2530 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2532 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2538 case Instruction::Store:
2539 case Instruction::Load:
2540 vectorizeMemoryInstruction(it, Legal);
2542 case Instruction::ZExt:
2543 case Instruction::SExt:
2544 case Instruction::FPToUI:
2545 case Instruction::FPToSI:
2546 case Instruction::FPExt:
2547 case Instruction::PtrToInt:
2548 case Instruction::IntToPtr:
2549 case Instruction::SIToFP:
2550 case Instruction::UIToFP:
2551 case Instruction::Trunc:
2552 case Instruction::FPTrunc:
2553 case Instruction::BitCast: {
2554 CastInst *CI = dyn_cast<CastInst>(it);
2555 DebugLocSetter SetDL(Builder, it);
2556 /// Optimize the special case where the source is the induction
2557 /// variable. Notice that we can only optimize the 'trunc' case
2558 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2559 /// c. other casts depend on pointer size.
2560 if (CI->getOperand(0) == OldInduction &&
2561 it->getOpcode() == Instruction::Trunc) {
2562 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2564 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2565 for (unsigned Part = 0; Part < UF; ++Part)
2566 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2569 /// Vectorize casts.
2570 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2572 VectorParts &A = getVectorValue(it->getOperand(0));
2573 for (unsigned Part = 0; Part < UF; ++Part)
2574 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2578 case Instruction::Call: {
2579 // Ignore dbg intrinsics.
2580 if (isa<DbgInfoIntrinsic>(it))
2583 DebugLocSetter SetDL(Builder, it);
2585 Module *M = BB->getParent()->getParent();
2586 CallInst *CI = cast<CallInst>(it);
2587 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2588 assert(ID && "Not an intrinsic call!");
2589 for (unsigned Part = 0; Part < UF; ++Part) {
2590 SmallVector<Value*, 4> Args;
2591 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2592 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2593 Args.push_back(Arg[Part]);
2595 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2596 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2597 Entry[Part] = Builder.CreateCall(F, Args);
2603 // All other instructions are unsupported. Scalarize them.
2604 scalarizeInstruction(it);
2607 }// end of for_each instr.
2610 void InnerLoopVectorizer::updateAnalysis() {
2611 // Forget the original basic block.
2612 SE->forgetLoop(OrigLoop);
2614 // Update the dominator tree information.
2615 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2616 "Entry does not dominate exit.");
2618 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2619 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2620 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2621 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2622 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2623 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2624 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2625 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2627 DEBUG(DT->verifyAnalysis());
2630 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2631 if (!EnableIfConversion)
2634 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2635 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2637 // Collect the blocks that need predication.
2638 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2639 BasicBlock *BB = LoopBlocks[i];
2641 // We don't support switch statements inside loops.
2642 if (!isa<BranchInst>(BB->getTerminator()))
2645 // We must be able to predicate all blocks that need to be predicated.
2646 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2650 // Check that we can actually speculate the hoistable loads.
2651 if (!LoadSpeculation.canHoistAllLoads())
2654 // We can if-convert this loop.
2658 bool LoopVectorizationLegality::canVectorize() {
2659 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2660 // be canonicalized.
2661 if (!TheLoop->getLoopPreheader())
2664 // We can only vectorize innermost loops.
2665 if (TheLoop->getSubLoopsVector().size())
2668 // We must have a single backedge.
2669 if (TheLoop->getNumBackEdges() != 1)
2672 // We must have a single exiting block.
2673 if (!TheLoop->getExitingBlock())
2676 unsigned NumBlocks = TheLoop->getNumBlocks();
2678 // Check if we can if-convert non single-bb loops.
2679 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2680 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2684 // We need to have a loop header.
2685 BasicBlock *Latch = TheLoop->getLoopLatch();
2686 DEBUG(dbgs() << "LV: Found a loop: " <<
2687 TheLoop->getHeader()->getName() << "\n");
2689 // ScalarEvolution needs to be able to find the exit count.
2690 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2691 if (ExitCount == SE->getCouldNotCompute()) {
2692 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2696 // Do not loop-vectorize loops with a tiny trip count.
2697 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2698 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2699 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2700 "This loop is not worth vectorizing.\n");
2704 // Check if we can vectorize the instructions and CFG in this loop.
2705 if (!canVectorizeInstrs()) {
2706 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2710 // Go over each instruction and look at memory deps.
2711 if (!canVectorizeMemory()) {
2712 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2716 // Collect all of the variables that remain uniform after vectorization.
2717 collectLoopUniforms();
2719 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2720 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2723 // Okay! We can vectorize. At this point we don't have any other mem analysis
2724 // which may limit our maximum vectorization factor, so just return true with
2729 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2730 if (Ty->isPointerTy())
2731 return DL.getIntPtrType(Ty->getContext());
2735 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2736 Ty0 = convertPointerToIntegerType(DL, Ty0);
2737 Ty1 = convertPointerToIntegerType(DL, Ty1);
2738 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2743 /// \brief Check that the instruction has outside loop users and is not an
2744 /// identified reduction variable.
2745 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2746 SmallPtrSet<Value *, 4> &Reductions) {
2747 // Reduction instructions are allowed to have exit users. All other
2748 // instructions must not have external users.
2749 if (!Reductions.count(Inst))
2750 //Check that all of the users of the loop are inside the BB.
2751 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2753 Instruction *U = cast<Instruction>(*I);
2754 // This user may be a reduction exit value.
2755 if (!TheLoop->contains(U)) {
2756 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2763 bool LoopVectorizationLegality::canVectorizeInstrs() {
2764 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2765 BasicBlock *Header = TheLoop->getHeader();
2767 // Look for the attribute signaling the absence of NaNs.
2768 Function &F = *Header->getParent();
2769 if (F.hasFnAttribute("no-nans-fp-math"))
2770 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2771 AttributeSet::FunctionIndex,
2772 "no-nans-fp-math").getValueAsString() == "true";
2774 // For each block in the loop.
2775 for (Loop::block_iterator bb = TheLoop->block_begin(),
2776 be = TheLoop->block_end(); bb != be; ++bb) {
2778 // Scan the instructions in the block and look for hazards.
2779 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2782 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2783 Type *PhiTy = Phi->getType();
2784 // Check that this PHI type is allowed.
2785 if (!PhiTy->isIntegerTy() &&
2786 !PhiTy->isFloatingPointTy() &&
2787 !PhiTy->isPointerTy()) {
2788 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2792 // If this PHINode is not in the header block, then we know that we
2793 // can convert it to select during if-conversion. No need to check if
2794 // the PHIs in this block are induction or reduction variables.
2795 if (*bb != Header) {
2796 // Check that this instruction has no outside users or is an
2797 // identified reduction value with an outside user.
2798 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2803 // We only allow if-converted PHIs with more than two incoming values.
2804 if (Phi->getNumIncomingValues() != 2) {
2805 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2809 // This is the value coming from the preheader.
2810 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2811 // Check if this is an induction variable.
2812 InductionKind IK = isInductionVariable(Phi);
2814 if (IK_NoInduction != IK) {
2815 // Get the widest type.
2817 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2819 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2821 // Int inductions are special because we only allow one IV.
2822 if (IK == IK_IntInduction) {
2823 // Use the phi node with the widest type as induction. Use the last
2824 // one if there are multiple (no good reason for doing this other
2825 // than it is expedient).
2826 if (!Induction || PhiTy == WidestIndTy)
2830 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2831 Inductions[Phi] = InductionInfo(StartValue, IK);
2835 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2836 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2839 if (AddReductionVar(Phi, RK_IntegerMult)) {
2840 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2843 if (AddReductionVar(Phi, RK_IntegerOr)) {
2844 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2847 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2848 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2851 if (AddReductionVar(Phi, RK_IntegerXor)) {
2852 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2855 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2856 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2859 if (AddReductionVar(Phi, RK_FloatMult)) {
2860 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2863 if (AddReductionVar(Phi, RK_FloatAdd)) {
2864 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2867 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2868 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2872 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2874 }// end of PHI handling
2876 // We still don't handle functions. However, we can ignore dbg intrinsic
2877 // calls and we do handle certain intrinsic and libm functions.
2878 CallInst *CI = dyn_cast<CallInst>(it);
2879 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2880 DEBUG(dbgs() << "LV: Found a call site.\n");
2884 // Check that the instruction return type is vectorizable.
2885 if (!VectorType::isValidElementType(it->getType()) &&
2886 !it->getType()->isVoidTy()) {
2887 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2891 // Check that the stored type is vectorizable.
2892 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2893 Type *T = ST->getValueOperand()->getType();
2894 if (!VectorType::isValidElementType(T))
2898 // Reduction instructions are allowed to have exit users.
2899 // All other instructions must not have external users.
2900 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2908 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2909 if (Inductions.empty())
2916 void LoopVectorizationLegality::collectLoopUniforms() {
2917 // We now know that the loop is vectorizable!
2918 // Collect variables that will remain uniform after vectorization.
2919 std::vector<Value*> Worklist;
2920 BasicBlock *Latch = TheLoop->getLoopLatch();
2922 // Start with the conditional branch and walk up the block.
2923 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2925 while (Worklist.size()) {
2926 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2927 Worklist.pop_back();
2929 // Look at instructions inside this loop.
2930 // Stop when reaching PHI nodes.
2931 // TODO: we need to follow values all over the loop, not only in this block.
2932 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2935 // This is a known uniform.
2938 // Insert all operands.
2939 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2943 /// \brief Analyses memory accesses in a loop.
2945 /// Checks whether run time pointer checks are needed and builds sets for data
2946 /// dependence checking.
2947 class AccessAnalysis {
2949 /// \brief Read or write access location.
2950 typedef std::pair<Value*, char> MemAccessInfo;
2952 /// \brief Set of potential dependent memory accesses.
2953 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2955 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2956 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2957 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2959 /// \brief Register a load and whether it is only read from.
2960 void addLoad(Value *Ptr, bool IsReadOnly) {
2961 Accesses.insert(std::make_pair(Ptr, false));
2963 ReadOnlyPtr.insert(Ptr);
2966 /// \brief Register a store.
2967 void addStore(Value *Ptr) {
2968 Accesses.insert(std::make_pair(Ptr, true));
2971 /// \brief Check whether we can check the pointers at runtime for
2972 /// non-intersection.
2973 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2974 unsigned &NumComparisons, ScalarEvolution *SE,
2977 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2978 /// and builds sets of dependent accesses.
2979 void buildDependenceSets() {
2980 // Process read-write pointers first.
2981 processMemAccesses(false);
2982 // Next, process read pointers.
2983 processMemAccesses(true);
2986 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
2988 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
2990 DenseSet<MemAccessInfo> &getDependenciesToCheck() { return CheckDeps; }
2993 typedef SetVector<MemAccessInfo> PtrAccessSet;
2994 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
2996 /// \brief Go over all memory access or only the deferred ones if
2997 /// \p UseDeferred is true and check whether runtime pointer checks are needed
2998 /// and build sets of dependency check candidates.
2999 void processMemAccesses(bool UseDeferred);
3001 /// Set of all accesses.
3002 PtrAccessSet Accesses;
3004 /// Set of access to check after all writes have been processed.
3005 PtrAccessSet DeferredAccesses;
3007 /// Map of pointers to last access encountered.
3008 UnderlyingObjToAccessMap ObjToLastAccess;
3010 /// Set of accesses that need a further dependence check.
3011 DenseSet<MemAccessInfo> CheckDeps;
3013 /// Set of pointers that are read only.
3014 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3016 /// Set of underlying objects already written to.
3017 SmallPtrSet<Value*, 16> WriteObjects;
3021 /// Sets of potentially dependent accesses - members of one set share an
3022 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3023 /// dependence check.
3024 DepCandidates &DepCands;
3026 bool AreAllWritesIdentified;
3027 bool AreAllReadsIdentified;
3028 bool IsRTCheckNeeded;
3031 /// \brief Check whether a pointer can participate in a runtime bounds check.
3032 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3033 const SCEV *PtrScev = SE->getSCEV(Ptr);
3034 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3038 return AR->isAffine();
3041 bool AccessAnalysis::canCheckPtrAtRT(
3042 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3043 unsigned &NumComparisons, ScalarEvolution *SE,
3045 // Find pointers with computable bounds. We are going to use this information
3046 // to place a runtime bound check.
3047 unsigned NumReadPtrChecks = 0;
3048 unsigned NumWritePtrChecks = 0;
3049 bool CanDoRT = true;
3051 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3052 // We assign consecutive id to access from different dependence sets.
3053 // Accesses within the same set don't need a runtime check.
3054 unsigned RunningDepId = 1;
3055 DenseMap<Value *, unsigned> DepSetId;
3057 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3059 const MemAccessInfo &Access = *AI;
3060 Value *Ptr = Access.first;
3061 bool IsWrite = Access.second;
3063 // Just add write checks if we have both.
3064 if (!IsWrite && Accesses.count(std::make_pair(Ptr, true)))
3068 ++NumWritePtrChecks;
3072 if (hasComputableBounds(SE, Ptr)) {
3073 // The id of the dependence set.
3076 if (IsDepCheckNeeded) {
3077 Value *Leader = DepCands.getLeaderValue(Access).first;
3078 unsigned &LeaderId = DepSetId[Leader];
3080 LeaderId = RunningDepId++;
3083 // Each access has its own dependence set.
3084 DepId = RunningDepId++;
3086 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3088 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3094 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3095 NumComparisons = 0; // Only one dependence set.
3097 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3098 NumWritePtrChecks - 1));
3102 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3103 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3106 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3107 // We process the set twice: first we process read-write pointers, last we
3108 // process read-only pointers. This allows us to skip dependence tests for
3109 // read-only pointers.
3111 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3112 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3113 const MemAccessInfo &Access = *AI;
3114 Value *Ptr = Access.first;
3115 bool IsWrite = Access.second;
3117 DepCands.insert(Access);
3119 // Memorize read-only pointers for later processing and skip them in the
3120 // first round (they need to be checked after we have seen all write
3121 // pointers). Note: we also mark pointer that are not consecutive as
3122 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3123 // second check for "!IsWrite".
3124 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3125 if (!UseDeferred && IsReadOnlyPtr) {
3126 DeferredAccesses.insert(Access);
3130 bool NeedDepCheck = false;
3131 // Check whether there is the possiblity of dependency because of underlying
3132 // objects being the same.
3133 typedef SmallVector<Value*, 16> ValueVector;
3134 ValueVector TempObjects;
3135 GetUnderlyingObjects(Ptr, TempObjects, DL);
3136 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3138 Value *UnderlyingObj = *UI;
3140 // If this is a write then it needs to be an identified object. If this a
3141 // read and all writes (so far) are identified function scope objects we
3142 // don't need an identified underlying object but only an Argument (the
3143 // next write is going to invalidate this assumption if it is
3145 // This is a micro-optimization for the case where all writes are
3146 // identified and we have one argument pointer.
3147 // Otherwise, we do need a runtime check.
3148 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3149 (!IsWrite && (!AreAllWritesIdentified ||
3150 !isa<Argument>(UnderlyingObj)) &&
3151 !isIdentifiedObject(UnderlyingObj))) {
3152 DEBUG(dbgs() << "LV: Found an unidentified " <<
3153 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3155 IsRTCheckNeeded = (IsRTCheckNeeded ||
3156 !isIdentifiedObject(UnderlyingObj) ||
3157 !AreAllReadsIdentified);
3160 AreAllWritesIdentified = false;
3162 AreAllReadsIdentified = false;
3165 // If this is a write - check other reads and writes for conflicts. If
3166 // this is a read only check other writes for conflicts (but only if there
3167 // is no other write to the ptr - this is an optimization to catch "a[i] =
3168 // a[i] + " without having to do a dependence check).
3169 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3170 NeedDepCheck = true;
3173 WriteObjects.insert(UnderlyingObj);
3175 // Create sets of pointers connected by shared underlying objects.
3176 UnderlyingObjToAccessMap::iterator Prev =
3177 ObjToLastAccess.find(UnderlyingObj);
3178 if (Prev != ObjToLastAccess.end())
3179 DepCands.unionSets(Access, Prev->second);
3181 ObjToLastAccess[UnderlyingObj] = Access;
3185 CheckDeps.insert(Access);
3189 /// \brief Checks memory dependences among accesses to the same underlying
3190 /// object to determine whether there vectorization is legal or not (and at
3191 /// which vectorization factor).
3193 /// This class works under the assumption that we already checked that memory
3194 /// locations with different underlying pointers are "must-not alias".
3195 /// We use the ScalarEvolution framework to symbolically evalutate access
3196 /// functions pairs. Since we currently don't restructure the loop we can rely
3197 /// on the program order of memory accesses to determine their safety.
3198 /// At the moment we will only deem accesses as safe for:
3199 /// * A negative constant distance assuming program order.
3201 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3202 /// a[i] = tmp; y = a[i];
3204 /// The latter case is safe because later checks guarantuee that there can't
3205 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3206 /// the same variable: a header phi can only be an induction or a reduction, a
3207 /// reduction can't have a memory sink, an induction can't have a memory
3208 /// source). This is important and must not be violated (or we have to
3209 /// resort to checking for cycles through memory).
3211 /// * A positive constant distance assuming program order that is bigger
3212 /// than the biggest memory access.
3214 /// tmp = a[i] OR b[i] = x
3215 /// a[i+2] = tmp y = b[i+2];
3217 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3219 /// * Zero distances and all accesses have the same size.
3221 class MemoryDepChecker {
3223 typedef std::pair<Value*, char> MemAccessInfo;
3225 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3226 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3228 /// \brief Register the location (instructions are given increasing numbers)
3229 /// of a write access.
3230 void addAccess(StoreInst *SI) {
3231 Value *Ptr = SI->getPointerOperand();
3232 Accesses[std::make_pair(Ptr, true)].push_back(AccessIdx);
3233 InstMap.push_back(SI);
3237 /// \brief Register the location (instructions are given increasing numbers)
3238 /// of a write access.
3239 void addAccess(LoadInst *LI) {
3240 Value *Ptr = LI->getPointerOperand();
3241 Accesses[std::make_pair(Ptr, false)].push_back(AccessIdx);
3242 InstMap.push_back(LI);
3246 /// \brief Check whether the dependencies between the accesses are safe.
3248 /// Only checks sets with elements in \p CheckDeps.
3249 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3250 DenseSet<MemAccessInfo> &CheckDeps);
3252 /// \brief The maximum number of bytes of a vector register we can vectorize
3253 /// the accesses safely with.
3254 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3257 ScalarEvolution *SE;
3259 const Loop *InnermostLoop;
3261 /// \brief Maps access locations (ptr, read/write) to program order.
3262 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3264 /// \brief Memory access instructions in program order.
3265 SmallVector<Instruction *, 16> InstMap;
3267 /// \brief The program order index to be used for the next instruction.
3270 // We can access this many bytes in parallel safely.
3271 unsigned MaxSafeDepDistBytes;
3273 /// \brief Check whether there is a plausible dependence between the two
3276 /// Access \p A must happen before \p B in program order. The two indices
3277 /// identify the index into the program order map.
3279 /// This function checks whether there is a plausible dependence (or the
3280 /// absence of such can't be proved) between the two accesses. If there is a
3281 /// plausible dependence but the dependence distance is bigger than one
3282 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3283 /// distance is smaller than any other distance encountered so far).
3284 /// Otherwise, this function returns true signaling a possible dependence.
3285 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3286 const MemAccessInfo &B, unsigned BIdx);
3288 /// \brief Check whether the data dependence could prevent store-load
3290 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3293 static bool isInBoundsGep(Value *Ptr) {
3294 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3295 return GEP->isInBounds();
3299 /// \brief Check whether the access through \p Ptr has a constant stride.
3300 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3302 const Type *PtrTy = Ptr->getType();
3303 assert(PtrTy->isPointerTy() && "Unexpected non ptr");
3305 // Make sure that the pointer does not point to aggregate types.
3306 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) {
3307 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr
3312 const SCEV *PtrScev = SE->getSCEV(Ptr);
3313 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3315 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3316 << *Ptr << " SCEV: " << *PtrScev << "\n");
3320 // The accesss function must stride over the innermost loop.
3321 if (Lp != AR->getLoop()) {
3322 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << *Ptr
3323 << " SCEV: " << *PtrScev << "\n");
3326 // The address calculation must not wrap. Otherwise, a dependence could be
3327 // inverted. An inbounds getelementptr that is a AddRec with a unit stride
3328 // cannot wrap per definition. The unit stride requirement is checked later.
3329 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3330 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3331 if (!IsNoWrapAddRec && !IsInBoundsGEP) {
3332 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3333 << *Ptr << " SCEV: " << *PtrScev << "\n");
3337 // Check the step is constant.
3338 const SCEV *Step = AR->getStepRecurrence(*SE);
3340 // Calculate the pointer stride and check if it is consecutive.
3341 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3343 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3344 " SCEV: " << *PtrScev << "\n");
3348 int64_t Size = DL->getTypeAllocSize(PtrTy->getPointerElementType());
3349 const APInt &APStepVal = C->getValue()->getValue();
3351 // Huge step value - give up.
3352 if (APStepVal.getBitWidth() > 64)
3355 int64_t StepVal = APStepVal.getSExtValue();
3358 int64_t Stride = StepVal / Size;
3359 int64_t Rem = StepVal % Size;
3363 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3364 // know we can't "wrap around the address space".
3365 if (!IsNoWrapAddRec && IsInBoundsGEP && Stride != 1 && Stride != -1)
3371 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3372 unsigned TypeByteSize) {
3373 // If loads occur at a distance that is not a multiple of a feasible vector
3374 // factor store-load forwarding does not take place.
3375 // Positive dependences might cause troubles because vectorizing them might
3376 // prevent store-load forwarding making vectorized code run a lot slower.
3377 // a[i] = a[i-3] ^ a[i-8];
3378 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3379 // hence on your typical architecture store-load forwarding does not take
3380 // place. Vectorizing in such cases does not make sense.
3381 // Store-load forwarding distance.
3382 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3383 // Maximum vector factor.
3384 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3385 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3386 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3388 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3390 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3391 MaxVFWithoutSLForwardIssues = (vf >>=1);
3396 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3397 DEBUG(dbgs() << "LV: Distance " << Distance <<
3398 " that could cause a store-load forwarding conflict\n");
3402 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3403 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3404 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3408 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3409 const MemAccessInfo &B, unsigned BIdx) {
3410 assert (AIdx < BIdx && "Must pass arguments in program order");
3412 Value *APtr = A.first;
3413 Value *BPtr = B.first;
3414 bool AIsWrite = A.second;
3415 bool BIsWrite = B.second;
3417 // Two reads are independent.
3418 if (!AIsWrite && !BIsWrite)
3421 const SCEV *AScev = SE->getSCEV(APtr);
3422 const SCEV *BScev = SE->getSCEV(BPtr);
3424 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3425 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3427 const SCEV *Src = AScev;
3428 const SCEV *Sink = BScev;
3430 // If the induction step is negative we have to invert source and sink of the
3432 if (StrideAPtr < 0) {
3435 std::swap(APtr, BPtr);
3436 std::swap(Src, Sink);
3437 std::swap(AIsWrite, BIsWrite);
3438 std::swap(AIdx, BIdx);
3439 std::swap(StrideAPtr, StrideBPtr);
3442 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3444 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3445 << "(Induction step: " << StrideAPtr << ")\n");
3446 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3447 << *InstMap[BIdx] << ": " << *Dist << "\n");
3449 // Need consecutive accesses. We don't want to vectorize
3450 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3451 // the address space.
3452 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3453 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3457 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3459 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3463 Type *ATy = APtr->getType()->getPointerElementType();
3464 Type *BTy = BPtr->getType()->getPointerElementType();
3465 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3467 // Negative distances are not plausible dependencies.
3468 const APInt &Val = C->getValue()->getValue();
3469 if (Val.isNegative()) {
3470 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3471 if (IsTrueDataDependence &&
3472 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3476 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3480 // Write to the same location with the same size.
3481 // Could be improved to assert type sizes are the same (i32 == float, etc).
3485 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3489 assert(Val.isStrictlyPositive() && "Expect a positive value");
3491 // Positive distance bigger than max vectorization factor.
3494 "LV: ReadWrite-Write positive dependency with different types");
3498 unsigned Distance = (unsigned) Val.getZExtValue();
3500 // Bail out early if passed-in parameters make vectorization not feasible.
3501 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3502 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3504 // The distance must be bigger than the size needed for a vectorized version
3505 // of the operation and the size of the vectorized operation must not be
3506 // bigger than the currrent maximum size.
3507 if (Distance < 2*TypeByteSize ||
3508 2*TypeByteSize > MaxSafeDepDistBytes ||
3509 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3510 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3511 << Val.getSExtValue() << "\n");
3515 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3516 Distance : MaxSafeDepDistBytes;
3518 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3519 if (IsTrueDataDependence &&
3520 couldPreventStoreLoadForward(Distance, TypeByteSize))
3523 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3524 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3530 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3531 DenseSet<MemAccessInfo> &CheckDeps) {
3533 MaxSafeDepDistBytes = -1U;
3534 while (!CheckDeps.empty()) {
3535 MemAccessInfo CurAccess = *CheckDeps.begin();
3537 // Get the relevant memory access set.
3538 EquivalenceClasses<MemAccessInfo>::iterator I =
3539 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3541 // Check accesses within this set.
3542 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3543 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3545 // Check every access pair.
3547 CheckDeps.erase(*AI);
3548 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3550 // Check every accessing instruction pair in program order.
3551 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3552 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3553 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3554 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3555 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3557 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3568 bool LoopVectorizationLegality::canVectorizeMemory() {
3570 typedef SmallVector<Value*, 16> ValueVector;
3571 typedef SmallPtrSet<Value*, 16> ValueSet;
3573 // Stores a pair of memory access location and whether the access is a store
3574 // (true) or a load (false).
3575 typedef std::pair<Value*, char> MemAccessInfo;
3576 typedef DenseSet<MemAccessInfo> PtrAccessSet;
3578 // Holds the Load and Store *instructions*.
3582 // Holds all the different accesses in the loop.
3583 unsigned NumReads = 0;
3584 unsigned NumReadWrites = 0;
3586 PtrRtCheck.Pointers.clear();
3587 PtrRtCheck.Need = false;
3589 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3590 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3593 for (Loop::block_iterator bb = TheLoop->block_begin(),
3594 be = TheLoop->block_end(); bb != be; ++bb) {
3596 // Scan the BB and collect legal loads and stores.
3597 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3600 // If this is a load, save it. If this instruction can read from memory
3601 // but is not a load, then we quit. Notice that we don't handle function
3602 // calls that read or write.
3603 if (it->mayReadFromMemory()) {
3604 LoadInst *Ld = dyn_cast<LoadInst>(it);
3605 if (!Ld) return false;
3606 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3607 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3610 Loads.push_back(Ld);
3611 DepChecker.addAccess(Ld);
3615 // Save 'store' instructions. Abort if other instructions write to memory.
3616 if (it->mayWriteToMemory()) {
3617 StoreInst *St = dyn_cast<StoreInst>(it);
3618 if (!St) return false;
3619 if (!St->isSimple() && !IsAnnotatedParallel) {
3620 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3623 Stores.push_back(St);
3624 DepChecker.addAccess(St);
3629 // Now we have two lists that hold the loads and the stores.
3630 // Next, we find the pointers that they use.
3632 // Check if we see any stores. If there are no stores, then we don't
3633 // care if the pointers are *restrict*.
3634 if (!Stores.size()) {
3635 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3639 AccessAnalysis::DepCandidates DependentAccesses;
3640 AccessAnalysis Accesses(DL, DependentAccesses);
3642 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3643 // multiple times on the same object. If the ptr is accessed twice, once
3644 // for read and once for write, it will only appear once (on the write
3645 // list). This is okay, since we are going to check for conflicts between
3646 // writes and between reads and writes, but not between reads and reads.
3649 ValueVector::iterator I, IE;
3650 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3651 StoreInst *ST = cast<StoreInst>(*I);
3652 Value* Ptr = ST->getPointerOperand();
3654 if (isUniform(Ptr)) {
3655 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3659 // If we did *not* see this pointer before, insert it to the read-write
3660 // list. At this phase it is only a 'write' list.
3661 if (Seen.insert(Ptr)) {
3663 Accesses.addStore(Ptr);
3667 if (IsAnnotatedParallel) {
3669 << "LV: A loop annotated parallel, ignore memory dependency "
3674 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3675 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3676 LoadInst *LD = cast<LoadInst>(*I);
3677 Value* Ptr = LD->getPointerOperand();
3678 // If we did *not* see this pointer before, insert it to the
3679 // read list. If we *did* see it before, then it is already in
3680 // the read-write list. This allows us to vectorize expressions
3681 // such as A[i] += x; Because the address of A[i] is a read-write
3682 // pointer. This only works if the index of A[i] is consecutive.
3683 // If the address of i is unknown (for example A[B[i]]) then we may
3684 // read a few words, modify, and write a few words, and some of the
3685 // words may be written to the same address.
3686 bool IsReadOnlyPtr = false;
3687 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3689 IsReadOnlyPtr = true;
3691 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3694 // If we write (or read-write) to a single destination and there are no
3695 // other reads in this loop then is it safe to vectorize.
3696 if (NumReadWrites == 1 && NumReads == 0) {
3697 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3701 // Build dependence sets and check whether we need a runtime pointer bounds
3703 Accesses.buildDependenceSets();
3704 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3706 // Find pointers with computable bounds. We are going to use this information
3707 // to place a runtime bound check.
3708 unsigned NumComparisons = 0;
3709 bool CanDoRT = false;
3711 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3714 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3715 " pointer comparisons.\n");
3717 // If we only have one set of dependences to check pointers among we don't
3718 // need a runtime check.
3719 if (NumComparisons == 0 && NeedRTCheck)
3720 NeedRTCheck = false;
3722 // Check that we did not collect too many pointers or found a unsizeable
3724 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3730 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3733 if (NeedRTCheck && !CanDoRT) {
3734 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3735 "the array bounds.\n");
3740 PtrRtCheck.Need = NeedRTCheck;
3742 bool CanVecMem = true;
3743 if (Accesses.isDependencyCheckNeeded()) {
3744 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3745 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3746 Accesses.getDependenciesToCheck());
3747 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3750 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3751 " need a runtime memory check.\n");
3756 static bool hasMultipleUsesOf(Instruction *I,
3757 SmallPtrSet<Instruction *, 8> &Insts) {
3758 unsigned NumUses = 0;
3759 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3760 if (Insts.count(dyn_cast<Instruction>(*Use)))
3769 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3770 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3771 if (!Set.count(dyn_cast<Instruction>(*Use)))
3776 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3777 ReductionKind Kind) {
3778 if (Phi->getNumIncomingValues() != 2)
3781 // Reduction variables are only found in the loop header block.
3782 if (Phi->getParent() != TheLoop->getHeader())
3785 // Obtain the reduction start value from the value that comes from the loop
3787 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3789 // ExitInstruction is the single value which is used outside the loop.
3790 // We only allow for a single reduction value to be used outside the loop.
3791 // This includes users of the reduction, variables (which form a cycle
3792 // which ends in the phi node).
3793 Instruction *ExitInstruction = 0;
3794 // Indicates that we found a reduction operation in our scan.
3795 bool FoundReduxOp = false;
3797 // We start with the PHI node and scan for all of the users of this
3798 // instruction. All users must be instructions that can be used as reduction
3799 // variables (such as ADD). We must have a single out-of-block user. The cycle
3800 // must include the original PHI.
3801 bool FoundStartPHI = false;
3803 // To recognize min/max patterns formed by a icmp select sequence, we store
3804 // the number of instruction we saw from the recognized min/max pattern,
3805 // to make sure we only see exactly the two instructions.
3806 unsigned NumCmpSelectPatternInst = 0;
3807 ReductionInstDesc ReduxDesc(false, 0);
3809 SmallPtrSet<Instruction *, 8> VisitedInsts;
3810 SmallVector<Instruction *, 8> Worklist;
3811 Worklist.push_back(Phi);
3812 VisitedInsts.insert(Phi);
3814 // A value in the reduction can be used:
3815 // - By the reduction:
3816 // - Reduction operation:
3817 // - One use of reduction value (safe).
3818 // - Multiple use of reduction value (not safe).
3820 // - All uses of the PHI must be the reduction (safe).
3821 // - Otherwise, not safe.
3822 // - By one instruction outside of the loop (safe).
3823 // - By further instructions outside of the loop (not safe).
3824 // - By an instruction that is not part of the reduction (not safe).
3826 // * An instruction type other than PHI or the reduction operation.
3827 // * A PHI in the header other than the initial PHI.
3828 while (!Worklist.empty()) {
3829 Instruction *Cur = Worklist.back();
3830 Worklist.pop_back();
3833 // If the instruction has no users then this is a broken chain and can't be
3834 // a reduction variable.
3835 if (Cur->use_empty())
3838 bool IsAPhi = isa<PHINode>(Cur);
3840 // A header PHI use other than the original PHI.
3841 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3844 // Reductions of instructions such as Div, and Sub is only possible if the
3845 // LHS is the reduction variable.
3846 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3847 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3848 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3851 // Any reduction instruction must be of one of the allowed kinds.
3852 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3853 if (!ReduxDesc.IsReduction)
3856 // A reduction operation must only have one use of the reduction value.
3857 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3858 hasMultipleUsesOf(Cur, VisitedInsts))
3861 // All inputs to a PHI node must be a reduction value.
3862 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3865 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3866 isa<SelectInst>(Cur)))
3867 ++NumCmpSelectPatternInst;
3868 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3869 isa<SelectInst>(Cur)))
3870 ++NumCmpSelectPatternInst;
3872 // Check whether we found a reduction operator.
3873 FoundReduxOp |= !IsAPhi;
3875 // Process users of current instruction. Push non PHI nodes after PHI nodes
3876 // onto the stack. This way we are going to have seen all inputs to PHI
3877 // nodes once we get to them.
3878 SmallVector<Instruction *, 8> NonPHIs;
3879 SmallVector<Instruction *, 8> PHIs;
3880 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3882 Instruction *Usr = cast<Instruction>(*UI);
3884 // Check if we found the exit user.
3885 BasicBlock *Parent = Usr->getParent();
3886 if (!TheLoop->contains(Parent)) {
3887 // Exit if you find multiple outside users.
3888 if (ExitInstruction != 0)
3890 ExitInstruction = Cur;
3894 // Process instructions only once (termination).
3895 if (VisitedInsts.insert(Usr)) {
3896 if (isa<PHINode>(Usr))
3897 PHIs.push_back(Usr);
3899 NonPHIs.push_back(Usr);
3901 // Remember that we completed the cycle.
3903 FoundStartPHI = true;
3905 Worklist.append(PHIs.begin(), PHIs.end());
3906 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3909 // This means we have seen one but not the other instruction of the
3910 // pattern or more than just a select and cmp.
3911 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3912 NumCmpSelectPatternInst != 2)
3915 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3918 // We found a reduction var if we have reached the original phi node and we
3919 // only have a single instruction with out-of-loop users.
3921 // This instruction is allowed to have out-of-loop users.
3922 AllowedExit.insert(ExitInstruction);
3924 // Save the description of this reduction variable.
3925 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3926 ReduxDesc.MinMaxKind);
3927 Reductions[Phi] = RD;
3928 // We've ended the cycle. This is a reduction variable if we have an
3929 // outside user and it has a binary op.
3934 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3935 /// pattern corresponding to a min(X, Y) or max(X, Y).
3936 LoopVectorizationLegality::ReductionInstDesc
3937 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3938 ReductionInstDesc &Prev) {
3940 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3941 "Expect a select instruction");
3942 Instruction *Cmp = 0;
3943 SelectInst *Select = 0;
3945 // We must handle the select(cmp()) as a single instruction. Advance to the
3947 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3948 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3949 return ReductionInstDesc(false, I);
3950 return ReductionInstDesc(Select, Prev.MinMaxKind);
3953 // Only handle single use cases for now.
3954 if (!(Select = dyn_cast<SelectInst>(I)))
3955 return ReductionInstDesc(false, I);
3956 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3957 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3958 return ReductionInstDesc(false, I);
3959 if (!Cmp->hasOneUse())
3960 return ReductionInstDesc(false, I);
3965 // Look for a min/max pattern.
3966 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3967 return ReductionInstDesc(Select, MRK_UIntMin);
3968 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3969 return ReductionInstDesc(Select, MRK_UIntMax);
3970 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3971 return ReductionInstDesc(Select, MRK_SIntMax);
3972 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3973 return ReductionInstDesc(Select, MRK_SIntMin);
3974 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3975 return ReductionInstDesc(Select, MRK_FloatMin);
3976 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3977 return ReductionInstDesc(Select, MRK_FloatMax);
3978 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3979 return ReductionInstDesc(Select, MRK_FloatMin);
3980 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3981 return ReductionInstDesc(Select, MRK_FloatMax);
3983 return ReductionInstDesc(false, I);
3986 LoopVectorizationLegality::ReductionInstDesc
3987 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3989 ReductionInstDesc &Prev) {
3990 bool FP = I->getType()->isFloatingPointTy();
3991 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3992 switch (I->getOpcode()) {
3994 return ReductionInstDesc(false, I);
3995 case Instruction::PHI:
3996 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3997 Kind != RK_FloatMinMax))
3998 return ReductionInstDesc(false, I);
3999 return ReductionInstDesc(I, Prev.MinMaxKind);
4000 case Instruction::Sub:
4001 case Instruction::Add:
4002 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4003 case Instruction::Mul:
4004 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4005 case Instruction::And:
4006 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4007 case Instruction::Or:
4008 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4009 case Instruction::Xor:
4010 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4011 case Instruction::FMul:
4012 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4013 case Instruction::FAdd:
4014 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4015 case Instruction::FCmp:
4016 case Instruction::ICmp:
4017 case Instruction::Select:
4018 if (Kind != RK_IntegerMinMax &&
4019 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4020 return ReductionInstDesc(false, I);
4021 return isMinMaxSelectCmpPattern(I, Prev);
4025 LoopVectorizationLegality::InductionKind
4026 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4027 Type *PhiTy = Phi->getType();
4028 // We only handle integer and pointer inductions variables.
4029 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4030 return IK_NoInduction;
4032 // Check that the PHI is consecutive.
4033 const SCEV *PhiScev = SE->getSCEV(Phi);
4034 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4036 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4037 return IK_NoInduction;
4039 const SCEV *Step = AR->getStepRecurrence(*SE);
4041 // Integer inductions need to have a stride of one.
4042 if (PhiTy->isIntegerTy()) {
4044 return IK_IntInduction;
4045 if (Step->isAllOnesValue())
4046 return IK_ReverseIntInduction;
4047 return IK_NoInduction;
4050 // Calculate the pointer stride and check if it is consecutive.
4051 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4053 return IK_NoInduction;
4055 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4056 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4057 if (C->getValue()->equalsInt(Size))
4058 return IK_PtrInduction;
4059 else if (C->getValue()->equalsInt(0 - Size))
4060 return IK_ReversePtrInduction;
4062 return IK_NoInduction;
4065 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4066 Value *In0 = const_cast<Value*>(V);
4067 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4071 return Inductions.count(PN);
4074 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4075 assert(TheLoop->contains(BB) && "Unknown block used");
4077 // Blocks that do not dominate the latch need predication.
4078 BasicBlock* Latch = TheLoop->getLoopLatch();
4079 return !DT->dominates(BB, Latch);
4082 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
4083 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4084 // We might be able to hoist the load.
4085 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
4088 // We don't predicate stores at the moment.
4089 if (it->mayWriteToMemory() || it->mayThrow())
4092 // The instructions below can trap.
4093 switch (it->getOpcode()) {
4095 case Instruction::UDiv:
4096 case Instruction::SDiv:
4097 case Instruction::URem:
4098 case Instruction::SRem:
4106 LoopVectorizationCostModel::VectorizationFactor
4107 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4109 // Width 1 means no vectorize
4110 VectorizationFactor Factor = { 1U, 0U };
4111 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4112 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4116 // Find the trip count.
4117 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4118 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4120 unsigned WidestType = getWidestType();
4121 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4122 unsigned MaxSafeDepDist = -1U;
4123 if (Legal->getMaxSafeDepDistBytes() != -1U)
4124 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4125 WidestRegister = WidestRegister < MaxSafeDepDist ? WidestRegister : MaxSafeDepDist;
4126 unsigned MaxVectorSize = WidestRegister / WidestType;
4127 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4128 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4130 if (MaxVectorSize == 0) {
4131 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4135 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4136 " into one vector!");
4138 unsigned VF = MaxVectorSize;
4140 // If we optimize the program for size, avoid creating the tail loop.
4142 // If we are unable to calculate the trip count then don't try to vectorize.
4144 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4148 // Find the maximum SIMD width that can fit within the trip count.
4149 VF = TC % MaxVectorSize;
4154 // If the trip count that we found modulo the vectorization factor is not
4155 // zero then we require a tail.
4157 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4163 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4164 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4166 Factor.Width = UserVF;
4170 float Cost = expectedCost(1);
4172 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4173 for (unsigned i=2; i <= VF; i*=2) {
4174 // Notice that the vector loop needs to be executed less times, so
4175 // we need to divide the cost of the vector loops by the width of
4176 // the vector elements.
4177 float VectorCost = expectedCost(i) / (float)i;
4178 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4179 (int)VectorCost << ".\n");
4180 if (VectorCost < Cost) {
4186 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4187 Factor.Width = Width;
4188 Factor.Cost = Width * Cost;
4192 unsigned LoopVectorizationCostModel::getWidestType() {
4193 unsigned MaxWidth = 8;
4196 for (Loop::block_iterator bb = TheLoop->block_begin(),
4197 be = TheLoop->block_end(); bb != be; ++bb) {
4198 BasicBlock *BB = *bb;
4200 // For each instruction in the loop.
4201 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4202 Type *T = it->getType();
4204 // Only examine Loads, Stores and PHINodes.
4205 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4208 // Examine PHI nodes that are reduction variables.
4209 if (PHINode *PN = dyn_cast<PHINode>(it))
4210 if (!Legal->getReductionVars()->count(PN))
4213 // Examine the stored values.
4214 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4215 T = ST->getValueOperand()->getType();
4217 // Ignore loaded pointer types and stored pointer types that are not
4218 // consecutive. However, we do want to take consecutive stores/loads of
4219 // pointer vectors into account.
4220 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4223 MaxWidth = std::max(MaxWidth,
4224 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4232 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4235 unsigned LoopCost) {
4237 // -- The unroll heuristics --
4238 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4239 // There are many micro-architectural considerations that we can't predict
4240 // at this level. For example frontend pressure (on decode or fetch) due to
4241 // code size, or the number and capabilities of the execution ports.
4243 // We use the following heuristics to select the unroll factor:
4244 // 1. If the code has reductions the we unroll in order to break the cross
4245 // iteration dependency.
4246 // 2. If the loop is really small then we unroll in order to reduce the loop
4248 // 3. We don't unroll if we think that we will spill registers to memory due
4249 // to the increased register pressure.
4251 // Use the user preference, unless 'auto' is selected.
4255 // When we optimize for size we don't unroll.
4259 // We used the distance for the unroll factor.
4260 if (Legal->getMaxSafeDepDistBytes() != -1U)
4263 // Do not unroll loops with a relatively small trip count.
4264 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4265 TheLoop->getLoopLatch());
4266 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4269 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4270 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4271 " vector registers\n");
4273 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4274 // We divide by these constants so assume that we have at least one
4275 // instruction that uses at least one register.
4276 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4277 R.NumInstructions = std::max(R.NumInstructions, 1U);
4279 // We calculate the unroll factor using the following formula.
4280 // Subtract the number of loop invariants from the number of available
4281 // registers. These registers are used by all of the unrolled instances.
4282 // Next, divide the remaining registers by the number of registers that is
4283 // required by the loop, in order to estimate how many parallel instances
4284 // fit without causing spills.
4285 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4287 // Clamp the unroll factor ranges to reasonable factors.
4288 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4290 // If we did not calculate the cost for VF (because the user selected the VF)
4291 // then we calculate the cost of VF here.
4293 LoopCost = expectedCost(VF);
4295 // Clamp the calculated UF to be between the 1 and the max unroll factor
4296 // that the target allows.
4297 if (UF > MaxUnrollSize)
4302 if (Legal->getReductionVars()->size()) {
4303 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4307 // We want to unroll tiny loops in order to reduce the loop overhead.
4308 // We assume that the cost overhead is 1 and we use the cost model
4309 // to estimate the cost of the loop and unroll until the cost of the
4310 // loop overhead is about 5% of the cost of the loop.
4311 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4312 if (LoopCost < 20) {
4313 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4314 unsigned NewUF = 20/LoopCost + 1;
4315 return std::min(NewUF, UF);
4318 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4322 LoopVectorizationCostModel::RegisterUsage
4323 LoopVectorizationCostModel::calculateRegisterUsage() {
4324 // This function calculates the register usage by measuring the highest number
4325 // of values that are alive at a single location. Obviously, this is a very
4326 // rough estimation. We scan the loop in a topological order in order and
4327 // assign a number to each instruction. We use RPO to ensure that defs are
4328 // met before their users. We assume that each instruction that has in-loop
4329 // users starts an interval. We record every time that an in-loop value is
4330 // used, so we have a list of the first and last occurrences of each
4331 // instruction. Next, we transpose this data structure into a multi map that
4332 // holds the list of intervals that *end* at a specific location. This multi
4333 // map allows us to perform a linear search. We scan the instructions linearly
4334 // and record each time that a new interval starts, by placing it in a set.
4335 // If we find this value in the multi-map then we remove it from the set.
4336 // The max register usage is the maximum size of the set.
4337 // We also search for instructions that are defined outside the loop, but are
4338 // used inside the loop. We need this number separately from the max-interval
4339 // usage number because when we unroll, loop-invariant values do not take
4341 LoopBlocksDFS DFS(TheLoop);
4345 R.NumInstructions = 0;
4347 // Each 'key' in the map opens a new interval. The values
4348 // of the map are the index of the 'last seen' usage of the
4349 // instruction that is the key.
4350 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4351 // Maps instruction to its index.
4352 DenseMap<unsigned, Instruction*> IdxToInstr;
4353 // Marks the end of each interval.
4354 IntervalMap EndPoint;
4355 // Saves the list of instruction indices that are used in the loop.
4356 SmallSet<Instruction*, 8> Ends;
4357 // Saves the list of values that are used in the loop but are
4358 // defined outside the loop, such as arguments and constants.
4359 SmallPtrSet<Value*, 8> LoopInvariants;
4362 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4363 be = DFS.endRPO(); bb != be; ++bb) {
4364 R.NumInstructions += (*bb)->size();
4365 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4367 Instruction *I = it;
4368 IdxToInstr[Index++] = I;
4370 // Save the end location of each USE.
4371 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4372 Value *U = I->getOperand(i);
4373 Instruction *Instr = dyn_cast<Instruction>(U);
4375 // Ignore non-instruction values such as arguments, constants, etc.
4376 if (!Instr) continue;
4378 // If this instruction is outside the loop then record it and continue.
4379 if (!TheLoop->contains(Instr)) {
4380 LoopInvariants.insert(Instr);
4384 // Overwrite previous end points.
4385 EndPoint[Instr] = Index;
4391 // Saves the list of intervals that end with the index in 'key'.
4392 typedef SmallVector<Instruction*, 2> InstrList;
4393 DenseMap<unsigned, InstrList> TransposeEnds;
4395 // Transpose the EndPoints to a list of values that end at each index.
4396 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4398 TransposeEnds[it->second].push_back(it->first);
4400 SmallSet<Instruction*, 8> OpenIntervals;
4401 unsigned MaxUsage = 0;
4404 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4405 for (unsigned int i = 0; i < Index; ++i) {
4406 Instruction *I = IdxToInstr[i];
4407 // Ignore instructions that are never used within the loop.
4408 if (!Ends.count(I)) continue;
4410 // Remove all of the instructions that end at this location.
4411 InstrList &List = TransposeEnds[i];
4412 for (unsigned int j=0, e = List.size(); j < e; ++j)
4413 OpenIntervals.erase(List[j]);
4415 // Count the number of live interals.
4416 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4418 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4419 OpenIntervals.size() <<"\n");
4421 // Add the current instruction to the list of open intervals.
4422 OpenIntervals.insert(I);
4425 unsigned Invariant = LoopInvariants.size();
4426 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4427 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4428 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4430 R.LoopInvariantRegs = Invariant;
4431 R.MaxLocalUsers = MaxUsage;
4435 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4439 for (Loop::block_iterator bb = TheLoop->block_begin(),
4440 be = TheLoop->block_end(); bb != be; ++bb) {
4441 unsigned BlockCost = 0;
4442 BasicBlock *BB = *bb;
4444 // For each instruction in the old loop.
4445 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4446 // Skip dbg intrinsics.
4447 if (isa<DbgInfoIntrinsic>(it))
4450 unsigned C = getInstructionCost(it, VF);
4452 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4453 VF << " For instruction: "<< *it << "\n");
4456 // We assume that if-converted blocks have a 50% chance of being executed.
4457 // When the code is scalar then some of the blocks are avoided due to CF.
4458 // When the code is vectorized we execute all code paths.
4459 if (Legal->blockNeedsPredication(*bb) && VF == 1)
4469 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4470 // If we know that this instruction will remain uniform, check the cost of
4471 // the scalar version.
4472 if (Legal->isUniformAfterVectorization(I))
4475 Type *RetTy = I->getType();
4476 Type *VectorTy = ToVectorTy(RetTy, VF);
4478 // TODO: We need to estimate the cost of intrinsic calls.
4479 switch (I->getOpcode()) {
4480 case Instruction::GetElementPtr:
4481 // We mark this instruction as zero-cost because the cost of GEPs in
4482 // vectorized code depends on whether the corresponding memory instruction
4483 // is scalarized or not. Therefore, we handle GEPs with the memory
4484 // instruction cost.
4486 case Instruction::Br: {
4487 return TTI.getCFInstrCost(I->getOpcode());
4489 case Instruction::PHI:
4490 //TODO: IF-converted IFs become selects.
4492 case Instruction::Add:
4493 case Instruction::FAdd:
4494 case Instruction::Sub:
4495 case Instruction::FSub:
4496 case Instruction::Mul:
4497 case Instruction::FMul:
4498 case Instruction::UDiv:
4499 case Instruction::SDiv:
4500 case Instruction::FDiv:
4501 case Instruction::URem:
4502 case Instruction::SRem:
4503 case Instruction::FRem:
4504 case Instruction::Shl:
4505 case Instruction::LShr:
4506 case Instruction::AShr:
4507 case Instruction::And:
4508 case Instruction::Or:
4509 case Instruction::Xor: {
4510 // Certain instructions can be cheaper to vectorize if they have a constant
4511 // second vector operand. One example of this are shifts on x86.
4512 TargetTransformInfo::OperandValueKind Op1VK =
4513 TargetTransformInfo::OK_AnyValue;
4514 TargetTransformInfo::OperandValueKind Op2VK =
4515 TargetTransformInfo::OK_AnyValue;
4517 if (isa<ConstantInt>(I->getOperand(1)))
4518 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4520 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4522 case Instruction::Select: {
4523 SelectInst *SI = cast<SelectInst>(I);
4524 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4525 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4526 Type *CondTy = SI->getCondition()->getType();
4528 CondTy = VectorType::get(CondTy, VF);
4530 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4532 case Instruction::ICmp:
4533 case Instruction::FCmp: {
4534 Type *ValTy = I->getOperand(0)->getType();
4535 VectorTy = ToVectorTy(ValTy, VF);
4536 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4538 case Instruction::Store:
4539 case Instruction::Load: {
4540 StoreInst *SI = dyn_cast<StoreInst>(I);
4541 LoadInst *LI = dyn_cast<LoadInst>(I);
4542 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4544 VectorTy = ToVectorTy(ValTy, VF);
4546 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4547 unsigned AS = SI ? SI->getPointerAddressSpace() :
4548 LI->getPointerAddressSpace();
4549 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4550 // We add the cost of address computation here instead of with the gep
4551 // instruction because only here we know whether the operation is
4554 return TTI.getAddressComputationCost(VectorTy) +
4555 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4557 // Scalarized loads/stores.
4558 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4559 bool Reverse = ConsecutiveStride < 0;
4560 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4561 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4562 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4564 // The cost of extracting from the value vector and pointer vector.
4565 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4566 for (unsigned i = 0; i < VF; ++i) {
4567 // The cost of extracting the pointer operand.
4568 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4569 // In case of STORE, the cost of ExtractElement from the vector.
4570 // In case of LOAD, the cost of InsertElement into the returned
4572 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4573 Instruction::InsertElement,
4577 // The cost of the scalar loads/stores.
4578 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
4579 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4584 // Wide load/stores.
4585 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4586 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4589 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4593 case Instruction::ZExt:
4594 case Instruction::SExt:
4595 case Instruction::FPToUI:
4596 case Instruction::FPToSI:
4597 case Instruction::FPExt:
4598 case Instruction::PtrToInt:
4599 case Instruction::IntToPtr:
4600 case Instruction::SIToFP:
4601 case Instruction::UIToFP:
4602 case Instruction::Trunc:
4603 case Instruction::FPTrunc:
4604 case Instruction::BitCast: {
4605 // We optimize the truncation of induction variable.
4606 // The cost of these is the same as the scalar operation.
4607 if (I->getOpcode() == Instruction::Trunc &&
4608 Legal->isInductionVariable(I->getOperand(0)))
4609 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4610 I->getOperand(0)->getType());
4612 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4613 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4615 case Instruction::Call: {
4616 CallInst *CI = cast<CallInst>(I);
4617 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4618 assert(ID && "Not an intrinsic call!");
4619 Type *RetTy = ToVectorTy(CI->getType(), VF);
4620 SmallVector<Type*, 4> Tys;
4621 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4622 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4623 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4626 // We are scalarizing the instruction. Return the cost of the scalar
4627 // instruction, plus the cost of insert and extract into vector
4628 // elements, times the vector width.
4631 if (!RetTy->isVoidTy() && VF != 1) {
4632 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4634 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4637 // The cost of inserting the results plus extracting each one of the
4639 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4642 // The cost of executing VF copies of the scalar instruction. This opcode
4643 // is unknown. Assume that it is the same as 'mul'.
4644 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4650 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4651 if (Scalar->isVoidTy() || VF == 1)
4653 return VectorType::get(Scalar, VF);
4656 char LoopVectorize::ID = 0;
4657 static const char lv_name[] = "Loop Vectorization";
4658 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4659 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4660 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4661 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4662 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4665 Pass *createLoopVectorizePass() {
4666 return new LoopVectorize();
4670 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4671 // Check for a store.
4672 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4673 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4675 // Check for a load.
4676 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4677 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;