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 Look for a meaningful debug location on the instruction or it's
331 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
336 if (I->getDebugLoc() != Empty)
339 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
340 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
341 if (OpInst->getDebugLoc() != Empty)
348 /// \brief Set the debug location in the builder using the debug location in the
350 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
351 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
352 B.SetCurrentDebugLocation(Inst->getDebugLoc());
354 B.SetCurrentDebugLocation(DebugLoc());
357 /// \brief Check if conditionally executed loads are hoistable.
359 /// This class has two functions: isHoistableLoad and canHoistAllLoads.
360 /// isHoistableLoad should be called on all load instructions that are executed
361 /// conditionally. After all conditional loads are processed, the client should
362 /// call canHoistAllLoads to determine if all of the conditional executed loads
363 /// have an unconditional memory access to the same memory address in the loop.
365 typedef SmallPtrSet<Value *, 8> MemorySet;
369 MemorySet CondLoadAddrSet;
372 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
374 /// \brief Check if the instruction is a load with a identifiable address.
375 bool isHoistableLoad(Instruction *L);
377 /// \brief Check if all of the conditional loads are hoistable because there
378 /// exists an unconditional memory access to the same address in the loop.
379 bool canHoistAllLoads();
382 bool LoadHoisting::isHoistableLoad(Instruction *L) {
383 LoadInst *LI = dyn_cast<LoadInst>(L);
387 CondLoadAddrSet.insert(LI->getPointerOperand());
391 static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
392 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
393 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
394 Set.insert(LI->getPointerOperand());
395 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
396 Set.insert(SI->getPointerOperand());
400 bool LoadHoisting::canHoistAllLoads() {
401 // No conditional loads.
402 if (CondLoadAddrSet.empty())
405 MemorySet UncondMemAccesses;
406 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
407 BasicBlock *LoopLatch = TheLoop->getLoopLatch();
409 // Iterate over the unconditional blocks and collect memory access addresses.
410 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
411 BasicBlock *BB = LoopBlocks[i];
413 // Ignore conditional blocks.
414 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
417 addMemAccesses(BB, UncondMemAccesses);
420 // And make sure there is a matching unconditional access for every
422 for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
423 ME = CondLoadAddrSet.end(); MI != ME; ++MI)
424 if (!UncondMemAccesses.count(*MI))
430 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
431 /// to what vectorization factor.
432 /// This class does not look at the profitability of vectorization, only the
433 /// legality. This class has two main kinds of checks:
434 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
435 /// will change the order of memory accesses in a way that will change the
436 /// correctness of the program.
437 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
438 /// checks for a number of different conditions, such as the availability of a
439 /// single induction variable, that all types are supported and vectorize-able,
440 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
441 /// This class is also used by InnerLoopVectorizer for identifying
442 /// induction variable and the different reduction variables.
443 class LoopVectorizationLegality {
445 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
446 DominatorTree *DT, TargetLibraryInfo *TLI)
447 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
448 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
449 MaxSafeDepDistBytes(-1U), LoadSpeculation(L, DT) {}
451 /// This enum represents the kinds of reductions that we support.
453 RK_NoReduction, ///< Not a reduction.
454 RK_IntegerAdd, ///< Sum of integers.
455 RK_IntegerMult, ///< Product of integers.
456 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
457 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
458 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
459 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
460 RK_FloatAdd, ///< Sum of floats.
461 RK_FloatMult, ///< Product of floats.
462 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
465 /// This enum represents the kinds of inductions that we support.
467 IK_NoInduction, ///< Not an induction variable.
468 IK_IntInduction, ///< Integer induction variable. Step = 1.
469 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
470 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
471 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
474 // This enum represents the kind of minmax reduction.
475 enum MinMaxReductionKind {
485 /// This POD struct holds information about reduction variables.
486 struct ReductionDescriptor {
487 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
488 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
490 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
491 MinMaxReductionKind MK)
492 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
494 // The starting value of the reduction.
495 // It does not have to be zero!
496 TrackingVH<Value> StartValue;
497 // The instruction who's value is used outside the loop.
498 Instruction *LoopExitInstr;
499 // The kind of the reduction.
501 // If this a min/max reduction the kind of reduction.
502 MinMaxReductionKind MinMaxKind;
505 /// This POD struct holds information about a potential reduction operation.
506 struct ReductionInstDesc {
507 ReductionInstDesc(bool IsRedux, Instruction *I) :
508 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
510 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
511 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
513 // Is this instruction a reduction candidate.
515 // The last instruction in a min/max pattern (select of the select(icmp())
516 // pattern), or the current reduction instruction otherwise.
517 Instruction *PatternLastInst;
518 // If this is a min/max pattern the comparison predicate.
519 MinMaxReductionKind MinMaxKind;
522 // This POD struct holds information about the memory runtime legality
523 // check that a group of pointers do not overlap.
524 struct RuntimePointerCheck {
525 RuntimePointerCheck() : Need(false) {}
527 /// Reset the state of the pointer runtime information.
535 /// Insert a pointer and calculate the start and end SCEVs.
536 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
539 /// This flag indicates if we need to add the runtime check.
541 /// Holds the pointers that we need to check.
542 SmallVector<TrackingVH<Value>, 2> Pointers;
543 /// Holds the pointer value at the beginning of the loop.
544 SmallVector<const SCEV*, 2> Starts;
545 /// Holds the pointer value at the end of the loop.
546 SmallVector<const SCEV*, 2> Ends;
547 /// Holds the information if this pointer is used for writing to memory.
548 SmallVector<bool, 2> IsWritePtr;
549 /// Holds the id of the set of pointers that could be dependent because of a
550 /// shared underlying object.
551 SmallVector<unsigned, 2> DependencySetId;
554 /// A POD for saving information about induction variables.
555 struct InductionInfo {
556 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
557 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
559 TrackingVH<Value> StartValue;
564 /// ReductionList contains the reduction descriptors for all
565 /// of the reductions that were found in the loop.
566 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
568 /// InductionList saves induction variables and maps them to the
569 /// induction descriptor.
570 typedef MapVector<PHINode*, InductionInfo> InductionList;
572 /// Returns true if it is legal to vectorize this loop.
573 /// This does not mean that it is profitable to vectorize this
574 /// loop, only that it is legal to do so.
577 /// Returns the Induction variable.
578 PHINode *getInduction() { return Induction; }
580 /// Returns the reduction variables found in the loop.
581 ReductionList *getReductionVars() { return &Reductions; }
583 /// Returns the induction variables found in the loop.
584 InductionList *getInductionVars() { return &Inductions; }
586 /// Returns the widest induction type.
587 Type *getWidestInductionType() { return WidestIndTy; }
589 /// Returns True if V is an induction variable in this loop.
590 bool isInductionVariable(const Value *V);
592 /// Return true if the block BB needs to be predicated in order for the loop
593 /// to be vectorized.
594 bool blockNeedsPredication(BasicBlock *BB);
596 /// Check if this pointer is consecutive when vectorizing. This happens
597 /// when the last index of the GEP is the induction variable, or that the
598 /// pointer itself is an induction variable.
599 /// This check allows us to vectorize A[idx] into a wide load/store.
601 /// 0 - Stride is unknown or non consecutive.
602 /// 1 - Address is consecutive.
603 /// -1 - Address is consecutive, and decreasing.
604 int isConsecutivePtr(Value *Ptr);
606 /// Returns true if the value V is uniform within the loop.
607 bool isUniform(Value *V);
609 /// Returns true if this instruction will remain scalar after vectorization.
610 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
612 /// Returns the information that we collected about runtime memory check.
613 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
615 /// This function returns the identity element (or neutral element) for
617 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
619 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
622 /// Check if a single basic block loop is vectorizable.
623 /// At this point we know that this is a loop with a constant trip count
624 /// and we only need to check individual instructions.
625 bool canVectorizeInstrs();
627 /// When we vectorize loops we may change the order in which
628 /// we read and write from memory. This method checks if it is
629 /// legal to vectorize the code, considering only memory constrains.
630 /// Returns true if the loop is vectorizable
631 bool canVectorizeMemory();
633 /// Return true if we can vectorize this loop using the IF-conversion
635 bool canVectorizeWithIfConvert();
637 /// Collect the variables that need to stay uniform after vectorization.
638 void collectLoopUniforms();
640 /// Return true if all of the instructions in the block can be speculatively
642 bool blockCanBePredicated(BasicBlock *BB);
644 /// Returns True, if 'Phi' is the kind of reduction variable for type
645 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
646 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
647 /// Returns a struct describing if the instruction 'I' can be a reduction
648 /// variable of type 'Kind'. If the reduction is a min/max pattern of
649 /// select(icmp()) this function advances the instruction pointer 'I' from the
650 /// compare instruction to the select instruction and stores this pointer in
651 /// 'PatternLastInst' member of the returned struct.
652 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
653 ReductionInstDesc &Desc);
654 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
655 /// pattern corresponding to a min(X, Y) or max(X, Y).
656 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
657 ReductionInstDesc &Prev);
658 /// Returns the induction kind of Phi. This function may return NoInduction
659 /// if the PHI is not an induction variable.
660 InductionKind isInductionVariable(PHINode *Phi);
662 /// The loop that we evaluate.
666 /// DataLayout analysis.
670 /// Target Library Info.
671 TargetLibraryInfo *TLI;
673 // --- vectorization state --- //
675 /// Holds the integer induction variable. This is the counter of the
678 /// Holds the reduction variables.
679 ReductionList Reductions;
680 /// Holds all of the induction variables that we found in the loop.
681 /// Notice that inductions don't need to start at zero and that induction
682 /// variables can be pointers.
683 InductionList Inductions;
684 /// Holds the widest induction type encountered.
687 /// Allowed outside users. This holds the reduction
688 /// vars which can be accessed from outside the loop.
689 SmallPtrSet<Value*, 4> AllowedExit;
690 /// This set holds the variables which are known to be uniform after
692 SmallPtrSet<Instruction*, 4> Uniforms;
693 /// We need to check that all of the pointers in this list are disjoint
695 RuntimePointerCheck PtrRtCheck;
696 /// Can we assume the absence of NaNs.
697 bool HasFunNoNaNAttr;
699 unsigned MaxSafeDepDistBytes;
701 /// Utility to determine whether loads can be speculated.
702 LoadHoisting LoadSpeculation;
705 /// LoopVectorizationCostModel - estimates the expected speedups due to
707 /// In many cases vectorization is not profitable. This can happen because of
708 /// a number of reasons. In this class we mainly attempt to predict the
709 /// expected speedup/slowdowns due to the supported instruction set. We use the
710 /// TargetTransformInfo to query the different backends for the cost of
711 /// different operations.
712 class LoopVectorizationCostModel {
714 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
715 LoopVectorizationLegality *Legal,
716 const TargetTransformInfo &TTI,
717 DataLayout *DL, const TargetLibraryInfo *TLI)
718 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
720 /// Information about vectorization costs
721 struct VectorizationFactor {
722 unsigned Width; // Vector width with best cost
723 unsigned Cost; // Cost of the loop with that width
725 /// \return The most profitable vectorization factor and the cost of that VF.
726 /// This method checks every power of two up to VF. If UserVF is not ZERO
727 /// then this vectorization factor will be selected if vectorization is
729 VectorizationFactor selectVectorizationFactor(bool OptForSize,
732 /// \return The size (in bits) of the widest type in the code that
733 /// needs to be vectorized. We ignore values that remain scalar such as
734 /// 64 bit loop indices.
735 unsigned getWidestType();
737 /// \return The most profitable unroll factor.
738 /// If UserUF is non-zero then this method finds the best unroll-factor
739 /// based on register pressure and other parameters.
740 /// VF and LoopCost are the selected vectorization factor and the cost of the
742 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
745 /// \brief A struct that represents some properties of the register usage
747 struct RegisterUsage {
748 /// Holds the number of loop invariant values that are used in the loop.
749 unsigned LoopInvariantRegs;
750 /// Holds the maximum number of concurrent live intervals in the loop.
751 unsigned MaxLocalUsers;
752 /// Holds the number of instructions in the loop.
753 unsigned NumInstructions;
756 /// \return information about the register usage of the loop.
757 RegisterUsage calculateRegisterUsage();
760 /// Returns the expected execution cost. The unit of the cost does
761 /// not matter because we use the 'cost' units to compare different
762 /// vector widths. The cost that is returned is *not* normalized by
763 /// the factor width.
764 unsigned expectedCost(unsigned VF);
766 /// Returns the execution time cost of an instruction for a given vector
767 /// width. Vector width of one means scalar.
768 unsigned getInstructionCost(Instruction *I, unsigned VF);
770 /// A helper function for converting Scalar types to vector types.
771 /// If the incoming type is void, we return void. If the VF is 1, we return
773 static Type* ToVectorTy(Type *Scalar, unsigned VF);
775 /// Returns whether the instruction is a load or store and will be a emitted
776 /// as a vector operation.
777 bool isConsecutiveLoadOrStore(Instruction *I);
779 /// The loop that we evaluate.
783 /// Loop Info analysis.
785 /// Vectorization legality.
786 LoopVectorizationLegality *Legal;
787 /// Vector target information.
788 const TargetTransformInfo &TTI;
789 /// Target data layout information.
791 /// Target Library Info.
792 const TargetLibraryInfo *TLI;
795 /// Utility class for getting and setting loop vectorizer hints in the form
796 /// of loop metadata.
797 struct LoopVectorizeHints {
798 /// Vectorization width.
800 /// Vectorization unroll factor.
803 LoopVectorizeHints(const Loop *L)
804 : Width(VectorizationFactor)
805 , Unroll(VectorizationUnroll)
806 , LoopID(L->getLoopID()) {
808 // The command line options override any loop metadata except for when
809 // width == 1 which is used to indicate the loop is already vectorized.
810 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
811 Width = VectorizationFactor;
812 if (VectorizationUnroll.getNumOccurrences() > 0)
813 Unroll = VectorizationUnroll;
816 /// Return the loop vectorizer metadata prefix.
817 static StringRef Prefix() { return "llvm.vectorizer."; }
819 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
820 SmallVector<Value*, 2> Vals;
821 Vals.push_back(MDString::get(Context, Name));
822 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
823 return MDNode::get(Context, Vals);
826 /// Mark the loop L as already vectorized by setting the width to 1.
827 void setAlreadyVectorized(Loop *L) {
828 LLVMContext &Context = L->getHeader()->getContext();
832 // Create a new loop id with one more operand for the already_vectorized
833 // hint. If the loop already has a loop id then copy the existing operands.
834 SmallVector<Value*, 4> Vals(1);
836 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
837 Vals.push_back(LoopID->getOperand(i));
839 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
841 MDNode *NewLoopID = MDNode::get(Context, Vals);
842 // Set operand 0 to refer to the loop id itself.
843 NewLoopID->replaceOperandWith(0, NewLoopID);
845 L->setLoopID(NewLoopID);
847 LoopID->replaceAllUsesWith(NewLoopID);
855 /// Find hints specified in the loop metadata.
856 void getHints(const Loop *L) {
860 // First operand should refer to the loop id itself.
861 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
862 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
864 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
865 const MDString *S = 0;
866 SmallVector<Value*, 4> Args;
868 // The expected hint is either a MDString or a MDNode with the first
869 // operand a MDString.
870 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
871 if (!MD || MD->getNumOperands() == 0)
873 S = dyn_cast<MDString>(MD->getOperand(0));
874 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
875 Args.push_back(MD->getOperand(i));
877 S = dyn_cast<MDString>(LoopID->getOperand(i));
878 assert(Args.size() == 0 && "too many arguments for MDString");
884 // Check if the hint starts with the vectorizer prefix.
885 StringRef Hint = S->getString();
886 if (!Hint.startswith(Prefix()))
888 // Remove the prefix.
889 Hint = Hint.substr(Prefix().size(), StringRef::npos);
891 if (Args.size() == 1)
892 getHint(Hint, Args[0]);
896 // Check string hint with one operand.
897 void getHint(StringRef Hint, Value *Arg) {
898 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
900 unsigned Val = C->getZExtValue();
902 if (Hint == "width") {
903 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
904 "Invalid width metadata");
906 } else if (Hint == "unroll") {
907 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
908 "Invalid unroll metadata");
911 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
915 /// The LoopVectorize Pass.
916 struct LoopVectorize : public LoopPass {
917 /// Pass identification, replacement for typeid
920 explicit LoopVectorize() : LoopPass(ID) {
921 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
927 TargetTransformInfo *TTI;
929 TargetLibraryInfo *TLI;
931 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
932 // We only vectorize innermost loops.
936 SE = &getAnalysis<ScalarEvolution>();
937 DL = getAnalysisIfAvailable<DataLayout>();
938 LI = &getAnalysis<LoopInfo>();
939 TTI = &getAnalysis<TargetTransformInfo>();
940 DT = &getAnalysis<DominatorTree>();
941 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
944 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
948 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
949 L->getHeader()->getParent()->getName() << "\"\n");
951 LoopVectorizeHints Hints(L);
953 if (Hints.Width == 1) {
954 DEBUG(dbgs() << "LV: Not vectorizing.\n");
958 // Check if it is legal to vectorize the loop.
959 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
960 if (!LVL.canVectorize()) {
961 DEBUG(dbgs() << "LV: Not vectorizing.\n");
965 // Use the cost model.
966 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
968 // Check the function attributes to find out if this function should be
969 // optimized for size.
970 Function *F = L->getHeader()->getParent();
971 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
972 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
973 unsigned FnIndex = AttributeSet::FunctionIndex;
974 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
975 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
978 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
979 "attribute is used.\n");
983 // Select the optimal vectorization factor.
984 LoopVectorizationCostModel::VectorizationFactor VF;
985 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
986 // Select the unroll factor.
987 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
991 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
995 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
996 F->getParent()->getModuleIdentifier()<<"\n");
997 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
999 // If we decided that it is *legal* to vectorize the loop then do it.
1000 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1003 // Mark the loop as already vectorized to avoid vectorizing again.
1004 Hints.setAlreadyVectorized(L);
1006 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1010 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1011 LoopPass::getAnalysisUsage(AU);
1012 AU.addRequiredID(LoopSimplifyID);
1013 AU.addRequiredID(LCSSAID);
1014 AU.addRequired<DominatorTree>();
1015 AU.addRequired<LoopInfo>();
1016 AU.addRequired<ScalarEvolution>();
1017 AU.addRequired<TargetTransformInfo>();
1018 AU.addPreserved<LoopInfo>();
1019 AU.addPreserved<DominatorTree>();
1024 } // end anonymous namespace
1026 //===----------------------------------------------------------------------===//
1027 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1028 // LoopVectorizationCostModel.
1029 //===----------------------------------------------------------------------===//
1032 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1033 Loop *Lp, Value *Ptr,
1035 unsigned DepSetId) {
1036 const SCEV *Sc = SE->getSCEV(Ptr);
1037 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1038 assert(AR && "Invalid addrec expression");
1039 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1040 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1041 Pointers.push_back(Ptr);
1042 Starts.push_back(AR->getStart());
1043 Ends.push_back(ScEnd);
1044 IsWritePtr.push_back(WritePtr);
1045 DependencySetId.push_back(DepSetId);
1048 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1049 // Save the current insertion location.
1050 Instruction *Loc = Builder.GetInsertPoint();
1052 // We need to place the broadcast of invariant variables outside the loop.
1053 Instruction *Instr = dyn_cast<Instruction>(V);
1054 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1055 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1057 // Place the code for broadcasting invariant variables in the new preheader.
1059 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1061 // Broadcast the scalar into all locations in the vector.
1062 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1064 // Restore the builder insertion point.
1066 Builder.SetInsertPoint(Loc);
1071 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1073 assert(Val->getType()->isVectorTy() && "Must be a vector");
1074 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1075 "Elem must be an integer");
1076 // Create the types.
1077 Type *ITy = Val->getType()->getScalarType();
1078 VectorType *Ty = cast<VectorType>(Val->getType());
1079 int VLen = Ty->getNumElements();
1080 SmallVector<Constant*, 8> Indices;
1082 // Create a vector of consecutive numbers from zero to VF.
1083 for (int i = 0; i < VLen; ++i) {
1084 int64_t Idx = Negate ? (-i) : i;
1085 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1088 // Add the consecutive indices to the vector value.
1089 Constant *Cv = ConstantVector::get(Indices);
1090 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1091 return Builder.CreateAdd(Val, Cv, "induction");
1094 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1095 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1096 // Make sure that the pointer does not point to structs.
1097 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1100 // If this value is a pointer induction variable we know it is consecutive.
1101 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1102 if (Phi && Inductions.count(Phi)) {
1103 InductionInfo II = Inductions[Phi];
1104 if (IK_PtrInduction == II.IK)
1106 else if (IK_ReversePtrInduction == II.IK)
1110 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1114 unsigned NumOperands = Gep->getNumOperands();
1115 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1117 Value *GpPtr = Gep->getPointerOperand();
1118 // If this GEP value is a consecutive pointer induction variable and all of
1119 // the indices are constant then we know it is consecutive. We can
1120 Phi = dyn_cast<PHINode>(GpPtr);
1121 if (Phi && Inductions.count(Phi)) {
1123 // Make sure that the pointer does not point to structs.
1124 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1125 if (GepPtrType->getElementType()->isAggregateType())
1128 // Make sure that all of the index operands are loop invariant.
1129 for (unsigned i = 1; i < NumOperands; ++i)
1130 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1133 InductionInfo II = Inductions[Phi];
1134 if (IK_PtrInduction == II.IK)
1136 else if (IK_ReversePtrInduction == II.IK)
1140 // Check that all of the gep indices are uniform except for the last.
1141 for (unsigned i = 0; i < NumOperands - 1; ++i)
1142 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1145 // We can emit wide load/stores only if the last index is the induction
1147 const SCEV *Last = SE->getSCEV(LastIndex);
1148 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1149 const SCEV *Step = AR->getStepRecurrence(*SE);
1151 // The memory is consecutive because the last index is consecutive
1152 // and all other indices are loop invariant.
1155 if (Step->isAllOnesValue())
1162 bool LoopVectorizationLegality::isUniform(Value *V) {
1163 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1166 InnerLoopVectorizer::VectorParts&
1167 InnerLoopVectorizer::getVectorValue(Value *V) {
1168 assert(V != Induction && "The new induction variable should not be used.");
1169 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1171 // If we have this scalar in the map, return it.
1172 if (WidenMap.has(V))
1173 return WidenMap.get(V);
1175 // If this scalar is unknown, assume that it is a constant or that it is
1176 // loop invariant. Broadcast V and save the value for future uses.
1177 Value *B = getBroadcastInstrs(V);
1178 return WidenMap.splat(V, B);
1181 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1182 assert(Vec->getType()->isVectorTy() && "Invalid type");
1183 SmallVector<Constant*, 8> ShuffleMask;
1184 for (unsigned i = 0; i < VF; ++i)
1185 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1187 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1188 ConstantVector::get(ShuffleMask),
1193 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1194 LoopVectorizationLegality *Legal) {
1195 // Attempt to issue a wide load.
1196 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1197 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1199 assert((LI || SI) && "Invalid Load/Store instruction");
1201 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1202 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1203 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1204 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1205 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1206 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1207 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1209 if (ScalarAllocatedSize != VectorElementSize)
1210 return scalarizeInstruction(Instr);
1212 // If the pointer is loop invariant or if it is non consecutive,
1213 // scalarize the load.
1214 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1215 bool Reverse = ConsecutiveStride < 0;
1216 bool UniformLoad = LI && Legal->isUniform(Ptr);
1217 if (!ConsecutiveStride || UniformLoad)
1218 return scalarizeInstruction(Instr);
1220 Constant *Zero = Builder.getInt32(0);
1221 VectorParts &Entry = WidenMap.get(Instr);
1223 // Handle consecutive loads/stores.
1224 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1225 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1226 setDebugLocFromInst(Builder, Gep);
1227 Value *PtrOperand = Gep->getPointerOperand();
1228 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1229 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1231 // Create the new GEP with the new induction variable.
1232 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1233 Gep2->setOperand(0, FirstBasePtr);
1234 Gep2->setName("gep.indvar.base");
1235 Ptr = Builder.Insert(Gep2);
1237 setDebugLocFromInst(Builder, Gep);
1238 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1239 OrigLoop) && "Base ptr must be invariant");
1241 // The last index does not have to be the induction. It can be
1242 // consecutive and be a function of the index. For example A[I+1];
1243 unsigned NumOperands = Gep->getNumOperands();
1244 unsigned LastOperand = NumOperands - 1;
1245 // Create the new GEP with the new induction variable.
1246 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1248 for (unsigned i = 0; i < NumOperands; ++i) {
1249 Value *GepOperand = Gep->getOperand(i);
1250 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1252 // Update last index or loop invariant instruction anchored in loop.
1253 if (i == LastOperand ||
1254 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1255 assert((i == LastOperand ||
1256 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1257 "Must be last index or loop invariant");
1259 VectorParts &GEPParts = getVectorValue(GepOperand);
1260 Value *Index = GEPParts[0];
1261 Index = Builder.CreateExtractElement(Index, Zero);
1262 Gep2->setOperand(i, Index);
1263 Gep2->setName("gep.indvar.idx");
1266 Ptr = Builder.Insert(Gep2);
1268 // Use the induction element ptr.
1269 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1270 setDebugLocFromInst(Builder, Ptr);
1271 VectorParts &PtrVal = getVectorValue(Ptr);
1272 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1277 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1278 "We do not allow storing to uniform addresses");
1279 setDebugLocFromInst(Builder, SI);
1280 // We don't want to update the value in the map as it might be used in
1281 // another expression. So don't use a reference type for "StoredVal".
1282 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1284 for (unsigned Part = 0; Part < UF; ++Part) {
1285 // Calculate the pointer for the specific unroll-part.
1286 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1289 // If we store to reverse consecutive memory locations then we need
1290 // to reverse the order of elements in the stored value.
1291 StoredVal[Part] = reverseVector(StoredVal[Part]);
1292 // If the address is consecutive but reversed, then the
1293 // wide store needs to start at the last vector element.
1294 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1295 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1298 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1299 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1305 assert(LI && "Must have a load instruction");
1306 setDebugLocFromInst(Builder, LI);
1307 for (unsigned Part = 0; Part < UF; ++Part) {
1308 // Calculate the pointer for the specific unroll-part.
1309 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1312 // If the address is consecutive but reversed, then the
1313 // wide store needs to start at the last vector element.
1314 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1315 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1318 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1319 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1320 cast<LoadInst>(LI)->setAlignment(Alignment);
1321 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1325 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1326 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1327 // Holds vector parameters or scalars, in case of uniform vals.
1328 SmallVector<VectorParts, 4> Params;
1330 setDebugLocFromInst(Builder, Instr);
1332 // Find all of the vectorized parameters.
1333 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1334 Value *SrcOp = Instr->getOperand(op);
1336 // If we are accessing the old induction variable, use the new one.
1337 if (SrcOp == OldInduction) {
1338 Params.push_back(getVectorValue(SrcOp));
1342 // Try using previously calculated values.
1343 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1345 // If the src is an instruction that appeared earlier in the basic block
1346 // then it should already be vectorized.
1347 if (SrcInst && OrigLoop->contains(SrcInst)) {
1348 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1349 // The parameter is a vector value from earlier.
1350 Params.push_back(WidenMap.get(SrcInst));
1352 // The parameter is a scalar from outside the loop. Maybe even a constant.
1353 VectorParts Scalars;
1354 Scalars.append(UF, SrcOp);
1355 Params.push_back(Scalars);
1359 assert(Params.size() == Instr->getNumOperands() &&
1360 "Invalid number of operands");
1362 // Does this instruction return a value ?
1363 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1365 Value *UndefVec = IsVoidRetTy ? 0 :
1366 UndefValue::get(VectorType::get(Instr->getType(), VF));
1367 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1368 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1370 // For each vector unroll 'part':
1371 for (unsigned Part = 0; Part < UF; ++Part) {
1372 // For each scalar that we create:
1373 for (unsigned Width = 0; Width < VF; ++Width) {
1374 Instruction *Cloned = Instr->clone();
1376 Cloned->setName(Instr->getName() + ".cloned");
1377 // Replace the operands of the cloned instrucions with extracted scalars.
1378 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1379 Value *Op = Params[op][Part];
1380 // Param is a vector. Need to extract the right lane.
1381 if (Op->getType()->isVectorTy())
1382 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1383 Cloned->setOperand(op, Op);
1386 // Place the cloned scalar in the new loop.
1387 Builder.Insert(Cloned);
1389 // If the original scalar returns a value we need to place it in a vector
1390 // so that future users will be able to use it.
1392 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1393 Builder.getInt32(Width));
1399 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1401 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1402 Legal->getRuntimePointerCheck();
1404 if (!PtrRtCheck->Need)
1407 unsigned NumPointers = PtrRtCheck->Pointers.size();
1408 SmallVector<TrackingVH<Value> , 2> Starts;
1409 SmallVector<TrackingVH<Value> , 2> Ends;
1411 SCEVExpander Exp(*SE, "induction");
1413 // Use this type for pointer arithmetic.
1414 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1416 for (unsigned i = 0; i < NumPointers; ++i) {
1417 Value *Ptr = PtrRtCheck->Pointers[i];
1418 const SCEV *Sc = SE->getSCEV(Ptr);
1420 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1421 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1423 Starts.push_back(Ptr);
1424 Ends.push_back(Ptr);
1426 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1428 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1429 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1430 Starts.push_back(Start);
1431 Ends.push_back(End);
1435 IRBuilder<> ChkBuilder(Loc);
1436 // Our instructions might fold to a constant.
1437 Value *MemoryRuntimeCheck = 0;
1438 for (unsigned i = 0; i < NumPointers; ++i) {
1439 for (unsigned j = i+1; j < NumPointers; ++j) {
1440 // No need to check if two readonly pointers intersect.
1441 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1444 // Only need to check pointers between two different dependency sets.
1445 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1448 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1449 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1450 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1451 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1453 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1454 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1455 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1456 if (MemoryRuntimeCheck)
1457 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1459 MemoryRuntimeCheck = IsConflict;
1463 // We have to do this trickery because the IRBuilder might fold the check to a
1464 // constant expression in which case there is no Instruction anchored in a
1466 LLVMContext &Ctx = Loc->getContext();
1467 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1468 ConstantInt::getTrue(Ctx));
1469 ChkBuilder.Insert(Check, "memcheck.conflict");
1474 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1476 In this function we generate a new loop. The new loop will contain
1477 the vectorized instructions while the old loop will continue to run the
1480 [ ] <-- vector loop bypass (may consist of multiple blocks).
1483 | [ ] <-- vector pre header.
1487 | [ ]_| <-- vector loop.
1490 >[ ] <--- middle-block.
1493 | [ ] <--- new preheader.
1497 | [ ]_| <-- old scalar loop to handle remainder.
1500 >[ ] <-- exit block.
1504 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1505 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1506 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1507 assert(ExitBlock && "Must have an exit block");
1509 // Some loops have a single integer induction variable, while other loops
1510 // don't. One example is c++ iterators that often have multiple pointer
1511 // induction variables. In the code below we also support a case where we
1512 // don't have a single induction variable.
1513 OldInduction = Legal->getInduction();
1514 Type *IdxTy = Legal->getWidestInductionType();
1516 // Find the loop boundaries.
1517 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1518 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1520 // Get the total trip count from the count by adding 1.
1521 ExitCount = SE->getAddExpr(ExitCount,
1522 SE->getConstant(ExitCount->getType(), 1));
1524 // Expand the trip count and place the new instructions in the preheader.
1525 // Notice that the pre-header does not change, only the loop body.
1526 SCEVExpander Exp(*SE, "induction");
1528 // Count holds the overall loop count (N).
1529 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1530 BypassBlock->getTerminator());
1532 // The loop index does not have to start at Zero. Find the original start
1533 // value from the induction PHI node. If we don't have an induction variable
1534 // then we know that it starts at zero.
1535 Builder.SetInsertPoint(BypassBlock->getTerminator());
1536 Value *StartIdx = ExtendedIdx = OldInduction ?
1537 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1539 ConstantInt::get(IdxTy, 0);
1541 assert(BypassBlock && "Invalid loop structure");
1542 LoopBypassBlocks.push_back(BypassBlock);
1544 // Split the single block loop into the two loop structure described above.
1545 BasicBlock *VectorPH =
1546 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1547 BasicBlock *VecBody =
1548 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1549 BasicBlock *MiddleBlock =
1550 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1551 BasicBlock *ScalarPH =
1552 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1554 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1556 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1558 // Generate the induction variable.
1559 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1560 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1561 // The loop step is equal to the vectorization factor (num of SIMD elements)
1562 // times the unroll factor (num of SIMD instructions).
1563 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1565 // This is the IR builder that we use to add all of the logic for bypassing
1566 // the new vector loop.
1567 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1568 setDebugLocFromInst(BypassBuilder,
1569 getDebugLocFromInstOrOperands(OldInduction));
1571 // We may need to extend the index in case there is a type mismatch.
1572 // We know that the count starts at zero and does not overflow.
1573 if (Count->getType() != IdxTy) {
1574 // The exit count can be of pointer type. Convert it to the correct
1576 if (ExitCount->getType()->isPointerTy())
1577 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1579 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1582 // Add the start index to the loop count to get the new end index.
1583 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1585 // Now we need to generate the expression for N - (N % VF), which is
1586 // the part that the vectorized body will execute.
1587 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1588 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1589 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1590 "end.idx.rnd.down");
1592 // Now, compare the new count to zero. If it is zero skip the vector loop and
1593 // jump to the scalar loop.
1594 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1597 BasicBlock *LastBypassBlock = BypassBlock;
1599 // Generate the code that checks in runtime if arrays overlap. We put the
1600 // checks into a separate block to make the more common case of few elements
1602 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1603 BypassBlock->getTerminator());
1604 if (MemRuntimeCheck) {
1605 // Create a new block containing the memory check.
1606 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1608 LoopBypassBlocks.push_back(CheckBlock);
1610 // Replace the branch into the memory check block with a conditional branch
1611 // for the "few elements case".
1612 Instruction *OldTerm = BypassBlock->getTerminator();
1613 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1614 OldTerm->eraseFromParent();
1616 Cmp = MemRuntimeCheck;
1617 LastBypassBlock = CheckBlock;
1620 LastBypassBlock->getTerminator()->eraseFromParent();
1621 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1624 // We are going to resume the execution of the scalar loop.
1625 // Go over all of the induction variables that we found and fix the
1626 // PHIs that are left in the scalar version of the loop.
1627 // The starting values of PHI nodes depend on the counter of the last
1628 // iteration in the vectorized loop.
1629 // If we come from a bypass edge then we need to start from the original
1632 // This variable saves the new starting index for the scalar loop.
1633 PHINode *ResumeIndex = 0;
1634 LoopVectorizationLegality::InductionList::iterator I, E;
1635 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1636 // Set builder to point to last bypass block.
1637 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1638 for (I = List->begin(), E = List->end(); I != E; ++I) {
1639 PHINode *OrigPhi = I->first;
1640 LoopVectorizationLegality::InductionInfo II = I->second;
1642 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1643 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1644 MiddleBlock->getTerminator());
1645 // We might have extended the type of the induction variable but we need a
1646 // truncated version for the scalar loop.
1647 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1648 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1649 MiddleBlock->getTerminator()) : 0;
1651 Value *EndValue = 0;
1653 case LoopVectorizationLegality::IK_NoInduction:
1654 llvm_unreachable("Unknown induction");
1655 case LoopVectorizationLegality::IK_IntInduction: {
1656 // Handle the integer induction counter.
1657 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1659 // We have the canonical induction variable.
1660 if (OrigPhi == OldInduction) {
1661 // Create a truncated version of the resume value for the scalar loop,
1662 // we might have promoted the type to a larger width.
1664 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1665 // The new PHI merges the original incoming value, in case of a bypass,
1666 // or the value at the end of the vectorized loop.
1667 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1668 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1669 TruncResumeVal->addIncoming(EndValue, VecBody);
1671 // We know what the end value is.
1672 EndValue = IdxEndRoundDown;
1673 // We also know which PHI node holds it.
1674 ResumeIndex = ResumeVal;
1678 // Not the canonical induction variable - add the vector loop count to the
1680 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1681 II.StartValue->getType(),
1683 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1686 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1687 // Convert the CountRoundDown variable to the PHI size.
1688 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1689 II.StartValue->getType(),
1691 // Handle reverse integer induction counter.
1692 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1695 case LoopVectorizationLegality::IK_PtrInduction: {
1696 // For pointer induction variables, calculate the offset using
1698 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1702 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1703 // The value at the end of the loop for the reverse pointer is calculated
1704 // by creating a GEP with a negative index starting from the start value.
1705 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1706 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1708 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1714 // The new PHI merges the original incoming value, in case of a bypass,
1715 // or the value at the end of the vectorized loop.
1716 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1717 if (OrigPhi == OldInduction)
1718 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1720 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1722 ResumeVal->addIncoming(EndValue, VecBody);
1724 // Fix the scalar body counter (PHI node).
1725 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1726 // The old inductions phi node in the scalar body needs the truncated value.
1727 if (OrigPhi == OldInduction)
1728 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1730 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1733 // If we are generating a new induction variable then we also need to
1734 // generate the code that calculates the exit value. This value is not
1735 // simply the end of the counter because we may skip the vectorized body
1736 // in case of a runtime check.
1738 assert(!ResumeIndex && "Unexpected resume value found");
1739 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1740 MiddleBlock->getTerminator());
1741 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1742 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1743 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1746 // Make sure that we found the index where scalar loop needs to continue.
1747 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1748 "Invalid resume Index");
1750 // Add a check in the middle block to see if we have completed
1751 // all of the iterations in the first vector loop.
1752 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1753 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1754 ResumeIndex, "cmp.n",
1755 MiddleBlock->getTerminator());
1757 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1758 // Remove the old terminator.
1759 MiddleBlock->getTerminator()->eraseFromParent();
1761 // Create i+1 and fill the PHINode.
1762 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1763 Induction->addIncoming(StartIdx, VectorPH);
1764 Induction->addIncoming(NextIdx, VecBody);
1765 // Create the compare.
1766 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1767 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1769 // Now we have two terminators. Remove the old one from the block.
1770 VecBody->getTerminator()->eraseFromParent();
1772 // Get ready to start creating new instructions into the vectorized body.
1773 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1775 // Create and register the new vector loop.
1776 Loop* Lp = new Loop();
1777 Loop *ParentLoop = OrigLoop->getParentLoop();
1779 // Insert the new loop into the loop nest and register the new basic blocks.
1781 ParentLoop->addChildLoop(Lp);
1782 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1783 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1784 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1785 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1786 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1788 LI->addTopLevelLoop(Lp);
1791 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1794 LoopVectorPreHeader = VectorPH;
1795 LoopScalarPreHeader = ScalarPH;
1796 LoopMiddleBlock = MiddleBlock;
1797 LoopExitBlock = ExitBlock;
1798 LoopVectorBody = VecBody;
1799 LoopScalarBody = OldBasicBlock;
1802 /// This function returns the identity element (or neutral element) for
1803 /// the operation K.
1805 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1810 // Adding, Xoring, Oring zero to a number does not change it.
1811 return ConstantInt::get(Tp, 0);
1812 case RK_IntegerMult:
1813 // Multiplying a number by 1 does not change it.
1814 return ConstantInt::get(Tp, 1);
1816 // AND-ing a number with an all-1 value does not change it.
1817 return ConstantInt::get(Tp, -1, true);
1819 // Multiplying a number by 1 does not change it.
1820 return ConstantFP::get(Tp, 1.0L);
1822 // Adding zero to a number does not change it.
1823 return ConstantFP::get(Tp, 0.0L);
1825 llvm_unreachable("Unknown reduction kind");
1829 static Intrinsic::ID
1830 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1831 // If we have an intrinsic call, check if it is trivially vectorizable.
1832 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1833 switch (II->getIntrinsicID()) {
1834 case Intrinsic::sqrt:
1835 case Intrinsic::sin:
1836 case Intrinsic::cos:
1837 case Intrinsic::exp:
1838 case Intrinsic::exp2:
1839 case Intrinsic::log:
1840 case Intrinsic::log10:
1841 case Intrinsic::log2:
1842 case Intrinsic::fabs:
1843 case Intrinsic::floor:
1844 case Intrinsic::ceil:
1845 case Intrinsic::trunc:
1846 case Intrinsic::rint:
1847 case Intrinsic::nearbyint:
1848 case Intrinsic::pow:
1849 case Intrinsic::fma:
1850 case Intrinsic::fmuladd:
1851 return II->getIntrinsicID();
1853 return Intrinsic::not_intrinsic;
1858 return Intrinsic::not_intrinsic;
1861 Function *F = CI->getCalledFunction();
1862 // We're going to make assumptions on the semantics of the functions, check
1863 // that the target knows that it's available in this environment.
1864 if (!F || !TLI->getLibFunc(F->getName(), Func))
1865 return Intrinsic::not_intrinsic;
1867 // Otherwise check if we have a call to a function that can be turned into a
1868 // vector intrinsic.
1875 return Intrinsic::sin;
1879 return Intrinsic::cos;
1883 return Intrinsic::exp;
1885 case LibFunc::exp2f:
1886 case LibFunc::exp2l:
1887 return Intrinsic::exp2;
1891 return Intrinsic::log;
1892 case LibFunc::log10:
1893 case LibFunc::log10f:
1894 case LibFunc::log10l:
1895 return Intrinsic::log10;
1897 case LibFunc::log2f:
1898 case LibFunc::log2l:
1899 return Intrinsic::log2;
1901 case LibFunc::fabsf:
1902 case LibFunc::fabsl:
1903 return Intrinsic::fabs;
1904 case LibFunc::floor:
1905 case LibFunc::floorf:
1906 case LibFunc::floorl:
1907 return Intrinsic::floor;
1909 case LibFunc::ceilf:
1910 case LibFunc::ceill:
1911 return Intrinsic::ceil;
1912 case LibFunc::trunc:
1913 case LibFunc::truncf:
1914 case LibFunc::truncl:
1915 return Intrinsic::trunc;
1917 case LibFunc::rintf:
1918 case LibFunc::rintl:
1919 return Intrinsic::rint;
1920 case LibFunc::nearbyint:
1921 case LibFunc::nearbyintf:
1922 case LibFunc::nearbyintl:
1923 return Intrinsic::nearbyint;
1927 return Intrinsic::pow;
1930 return Intrinsic::not_intrinsic;
1933 /// This function translates the reduction kind to an LLVM binary operator.
1935 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1937 case LoopVectorizationLegality::RK_IntegerAdd:
1938 return Instruction::Add;
1939 case LoopVectorizationLegality::RK_IntegerMult:
1940 return Instruction::Mul;
1941 case LoopVectorizationLegality::RK_IntegerOr:
1942 return Instruction::Or;
1943 case LoopVectorizationLegality::RK_IntegerAnd:
1944 return Instruction::And;
1945 case LoopVectorizationLegality::RK_IntegerXor:
1946 return Instruction::Xor;
1947 case LoopVectorizationLegality::RK_FloatMult:
1948 return Instruction::FMul;
1949 case LoopVectorizationLegality::RK_FloatAdd:
1950 return Instruction::FAdd;
1951 case LoopVectorizationLegality::RK_IntegerMinMax:
1952 return Instruction::ICmp;
1953 case LoopVectorizationLegality::RK_FloatMinMax:
1954 return Instruction::FCmp;
1956 llvm_unreachable("Unknown reduction operation");
1960 Value *createMinMaxOp(IRBuilder<> &Builder,
1961 LoopVectorizationLegality::MinMaxReductionKind RK,
1964 CmpInst::Predicate P = CmpInst::ICMP_NE;
1967 llvm_unreachable("Unknown min/max reduction kind");
1968 case LoopVectorizationLegality::MRK_UIntMin:
1969 P = CmpInst::ICMP_ULT;
1971 case LoopVectorizationLegality::MRK_UIntMax:
1972 P = CmpInst::ICMP_UGT;
1974 case LoopVectorizationLegality::MRK_SIntMin:
1975 P = CmpInst::ICMP_SLT;
1977 case LoopVectorizationLegality::MRK_SIntMax:
1978 P = CmpInst::ICMP_SGT;
1980 case LoopVectorizationLegality::MRK_FloatMin:
1981 P = CmpInst::FCMP_OLT;
1983 case LoopVectorizationLegality::MRK_FloatMax:
1984 P = CmpInst::FCMP_OGT;
1989 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1990 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1992 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1994 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1999 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2000 //===------------------------------------------------===//
2002 // Notice: any optimization or new instruction that go
2003 // into the code below should be also be implemented in
2006 //===------------------------------------------------===//
2007 Constant *Zero = Builder.getInt32(0);
2009 // In order to support reduction variables we need to be able to vectorize
2010 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2011 // stages. First, we create a new vector PHI node with no incoming edges.
2012 // We use this value when we vectorize all of the instructions that use the
2013 // PHI. Next, after all of the instructions in the block are complete we
2014 // add the new incoming edges to the PHI. At this point all of the
2015 // instructions in the basic block are vectorized, so we can use them to
2016 // construct the PHI.
2017 PhiVector RdxPHIsToFix;
2019 // Scan the loop in a topological order to ensure that defs are vectorized
2021 LoopBlocksDFS DFS(OrigLoop);
2024 // Vectorize all of the blocks in the original loop.
2025 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2026 be = DFS.endRPO(); bb != be; ++bb)
2027 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2029 // At this point every instruction in the original loop is widened to
2030 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2031 // that we vectorized. The PHI nodes are currently empty because we did
2032 // not want to introduce cycles. Notice that the remaining PHI nodes
2033 // that we need to fix are reduction variables.
2035 // Create the 'reduced' values for each of the induction vars.
2036 // The reduced values are the vector values that we scalarize and combine
2037 // after the loop is finished.
2038 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2040 PHINode *RdxPhi = *it;
2041 assert(RdxPhi && "Unable to recover vectorized PHI");
2043 // Find the reduction variable descriptor.
2044 assert(Legal->getReductionVars()->count(RdxPhi) &&
2045 "Unable to find the reduction variable");
2046 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2047 (*Legal->getReductionVars())[RdxPhi];
2049 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2051 // We need to generate a reduction vector from the incoming scalar.
2052 // To do so, we need to generate the 'identity' vector and overide
2053 // one of the elements with the incoming scalar reduction. We need
2054 // to do it in the vector-loop preheader.
2055 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2057 // This is the vector-clone of the value that leaves the loop.
2058 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2059 Type *VecTy = VectorExit[0]->getType();
2061 // Find the reduction identity variable. Zero for addition, or, xor,
2062 // one for multiplication, -1 for And.
2065 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2066 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2067 // MinMax reduction have the start value as their identify.
2068 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2072 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2073 VecTy->getScalarType());
2074 Identity = ConstantVector::getSplat(VF, Iden);
2076 // This vector is the Identity vector where the first element is the
2077 // incoming scalar reduction.
2078 VectorStart = Builder.CreateInsertElement(Identity,
2079 RdxDesc.StartValue, Zero);
2082 // Fix the vector-loop phi.
2083 // We created the induction variable so we know that the
2084 // preheader is the first entry.
2085 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2087 // Reductions do not have to start at zero. They can start with
2088 // any loop invariant values.
2089 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2090 BasicBlock *Latch = OrigLoop->getLoopLatch();
2091 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2092 VectorParts &Val = getVectorValue(LoopVal);
2093 for (unsigned part = 0; part < UF; ++part) {
2094 // Make sure to add the reduction stat value only to the
2095 // first unroll part.
2096 Value *StartVal = (part == 0) ? VectorStart : Identity;
2097 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2098 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2101 // Before each round, move the insertion point right between
2102 // the PHIs and the values we are going to write.
2103 // This allows us to write both PHINodes and the extractelement
2105 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2107 VectorParts RdxParts;
2108 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2109 for (unsigned part = 0; part < UF; ++part) {
2110 // This PHINode contains the vectorized reduction variable, or
2111 // the initial value vector, if we bypass the vector loop.
2112 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2113 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2114 Value *StartVal = (part == 0) ? VectorStart : Identity;
2115 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2116 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2117 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2118 RdxParts.push_back(NewPhi);
2121 // Reduce all of the unrolled parts into a single vector.
2122 Value *ReducedPartRdx = RdxParts[0];
2123 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2124 setDebugLocFromInst(Builder, ReducedPartRdx);
2125 for (unsigned part = 1; part < UF; ++part) {
2126 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2127 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2128 RdxParts[part], ReducedPartRdx,
2131 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2132 ReducedPartRdx, RdxParts[part]);
2135 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2136 // and vector ops, reducing the set of values being computed by half each
2138 assert(isPowerOf2_32(VF) &&
2139 "Reduction emission only supported for pow2 vectors!");
2140 Value *TmpVec = ReducedPartRdx;
2141 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2142 for (unsigned i = VF; i != 1; i >>= 1) {
2143 // Move the upper half of the vector to the lower half.
2144 for (unsigned j = 0; j != i/2; ++j)
2145 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2147 // Fill the rest of the mask with undef.
2148 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2149 UndefValue::get(Builder.getInt32Ty()));
2152 Builder.CreateShuffleVector(TmpVec,
2153 UndefValue::get(TmpVec->getType()),
2154 ConstantVector::get(ShuffleMask),
2157 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2158 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2161 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2164 // The result is in the first element of the vector.
2165 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2167 // Now, we need to fix the users of the reduction variable
2168 // inside and outside of the scalar remainder loop.
2169 // We know that the loop is in LCSSA form. We need to update the
2170 // PHI nodes in the exit blocks.
2171 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2172 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2173 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2174 if (!LCSSAPhi) continue;
2176 // All PHINodes need to have a single entry edge, or two if
2177 // we already fixed them.
2178 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2180 // We found our reduction value exit-PHI. Update it with the
2181 // incoming bypass edge.
2182 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2183 // Add an edge coming from the bypass.
2184 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2187 }// end of the LCSSA phi scan.
2189 // Fix the scalar loop reduction variable with the incoming reduction sum
2190 // from the vector body and from the backedge value.
2191 int IncomingEdgeBlockIdx =
2192 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2193 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2194 // Pick the other block.
2195 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2196 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2197 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2198 }// end of for each redux variable.
2200 // The Loop exit block may have single value PHI nodes where the incoming
2201 // value is 'undef'. While vectorizing we only handled real values that
2202 // were defined inside the loop. Here we handle the 'undef case'.
2204 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2205 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2206 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2207 if (!LCSSAPhi) continue;
2208 if (LCSSAPhi->getNumIncomingValues() == 1)
2209 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2214 InnerLoopVectorizer::VectorParts
2215 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2216 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2219 // Look for cached value.
2220 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2221 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2222 if (ECEntryIt != MaskCache.end())
2223 return ECEntryIt->second;
2225 VectorParts SrcMask = createBlockInMask(Src);
2227 // The terminator has to be a branch inst!
2228 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2229 assert(BI && "Unexpected terminator found");
2231 if (BI->isConditional()) {
2232 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2234 if (BI->getSuccessor(0) != Dst)
2235 for (unsigned part = 0; part < UF; ++part)
2236 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2238 for (unsigned part = 0; part < UF; ++part)
2239 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2241 MaskCache[Edge] = EdgeMask;
2245 MaskCache[Edge] = SrcMask;
2249 InnerLoopVectorizer::VectorParts
2250 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2251 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2253 // Loop incoming mask is all-one.
2254 if (OrigLoop->getHeader() == BB) {
2255 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2256 return getVectorValue(C);
2259 // This is the block mask. We OR all incoming edges, and with zero.
2260 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2261 VectorParts BlockMask = getVectorValue(Zero);
2264 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2265 VectorParts EM = createEdgeMask(*it, BB);
2266 for (unsigned part = 0; part < UF; ++part)
2267 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2274 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2275 BasicBlock *BB, PhiVector *PV) {
2276 // For each instruction in the old loop.
2277 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2278 VectorParts &Entry = WidenMap.get(it);
2279 switch (it->getOpcode()) {
2280 case Instruction::Br:
2281 // Nothing to do for PHIs and BR, since we already took care of the
2282 // loop control flow instructions.
2284 case Instruction::PHI:{
2285 PHINode* P = cast<PHINode>(it);
2286 // Handle reduction variables:
2287 if (Legal->getReductionVars()->count(P)) {
2288 for (unsigned part = 0; part < UF; ++part) {
2289 // This is phase one of vectorizing PHIs.
2290 Type *VecTy = VectorType::get(it->getType(), VF);
2291 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2292 LoopVectorBody-> getFirstInsertionPt());
2298 setDebugLocFromInst(Builder, P);
2299 // Check for PHI nodes that are lowered to vector selects.
2300 if (P->getParent() != OrigLoop->getHeader()) {
2301 // We know that all PHIs in non header blocks are converted into
2302 // selects, so we don't have to worry about the insertion order and we
2303 // can just use the builder.
2304 // At this point we generate the predication tree. There may be
2305 // duplications since this is a simple recursive scan, but future
2306 // optimizations will clean it up.
2308 unsigned NumIncoming = P->getNumIncomingValues();
2310 // Generate a sequence of selects of the form:
2311 // SELECT(Mask3, In3,
2312 // SELECT(Mask2, In2,
2314 for (unsigned In = 0; In < NumIncoming; In++) {
2315 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2317 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2319 for (unsigned part = 0; part < UF; ++part) {
2320 // We might have single edge PHIs (blocks) - use an identity
2321 // 'select' for the first PHI operand.
2323 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2326 // Select between the current value and the previous incoming edge
2327 // based on the incoming mask.
2328 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2329 Entry[part], "predphi");
2335 // This PHINode must be an induction variable.
2336 // Make sure that we know about it.
2337 assert(Legal->getInductionVars()->count(P) &&
2338 "Not an induction variable");
2340 LoopVectorizationLegality::InductionInfo II =
2341 Legal->getInductionVars()->lookup(P);
2344 case LoopVectorizationLegality::IK_NoInduction:
2345 llvm_unreachable("Unknown induction");
2346 case LoopVectorizationLegality::IK_IntInduction: {
2347 assert(P->getType() == II.StartValue->getType() && "Types must match");
2348 Type *PhiTy = P->getType();
2350 if (P == OldInduction) {
2351 // Handle the canonical induction variable. We might have had to
2353 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2355 // Handle other induction variables that are now based on the
2357 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2359 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2360 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2363 Broadcasted = getBroadcastInstrs(Broadcasted);
2364 // After broadcasting the induction variable we need to make the vector
2365 // consecutive by adding 0, 1, 2, etc.
2366 for (unsigned part = 0; part < UF; ++part)
2367 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2370 case LoopVectorizationLegality::IK_ReverseIntInduction:
2371 case LoopVectorizationLegality::IK_PtrInduction:
2372 case LoopVectorizationLegality::IK_ReversePtrInduction:
2373 // Handle reverse integer and pointer inductions.
2374 Value *StartIdx = ExtendedIdx;
2375 // This is the normalized GEP that starts counting at zero.
2376 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2379 // Handle the reverse integer induction variable case.
2380 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2381 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2382 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2384 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2387 // This is a new value so do not hoist it out.
2388 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2389 // After broadcasting the induction variable we need to make the
2390 // vector consecutive by adding ... -3, -2, -1, 0.
2391 for (unsigned part = 0; part < UF; ++part)
2392 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2397 // Handle the pointer induction variable case.
2398 assert(P->getType()->isPointerTy() && "Unexpected type.");
2400 // Is this a reverse induction ptr or a consecutive induction ptr.
2401 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2404 // This is the vector of results. Notice that we don't generate
2405 // vector geps because scalar geps result in better code.
2406 for (unsigned part = 0; part < UF; ++part) {
2407 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2408 for (unsigned int i = 0; i < VF; ++i) {
2409 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2410 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2413 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2415 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2417 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2419 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2420 Builder.getInt32(i),
2423 Entry[part] = VecVal;
2430 case Instruction::Add:
2431 case Instruction::FAdd:
2432 case Instruction::Sub:
2433 case Instruction::FSub:
2434 case Instruction::Mul:
2435 case Instruction::FMul:
2436 case Instruction::UDiv:
2437 case Instruction::SDiv:
2438 case Instruction::FDiv:
2439 case Instruction::URem:
2440 case Instruction::SRem:
2441 case Instruction::FRem:
2442 case Instruction::Shl:
2443 case Instruction::LShr:
2444 case Instruction::AShr:
2445 case Instruction::And:
2446 case Instruction::Or:
2447 case Instruction::Xor: {
2448 // Just widen binops.
2449 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2450 setDebugLocFromInst(Builder, BinOp);
2451 VectorParts &A = getVectorValue(it->getOperand(0));
2452 VectorParts &B = getVectorValue(it->getOperand(1));
2454 // Use this vector value for all users of the original instruction.
2455 for (unsigned Part = 0; Part < UF; ++Part) {
2456 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2458 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2459 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2460 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2461 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2462 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2464 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2465 VecOp->setIsExact(BinOp->isExact());
2471 case Instruction::Select: {
2473 // If the selector is loop invariant we can create a select
2474 // instruction with a scalar condition. Otherwise, use vector-select.
2475 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2477 setDebugLocFromInst(Builder, it);
2479 // The condition can be loop invariant but still defined inside the
2480 // loop. This means that we can't just use the original 'cond' value.
2481 // We have to take the 'vectorized' value and pick the first lane.
2482 // Instcombine will make this a no-op.
2483 VectorParts &Cond = getVectorValue(it->getOperand(0));
2484 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2485 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2486 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2487 Builder.getInt32(0));
2488 for (unsigned Part = 0; Part < UF; ++Part) {
2489 Entry[Part] = Builder.CreateSelect(
2490 InvariantCond ? ScalarCond : Cond[Part],
2497 case Instruction::ICmp:
2498 case Instruction::FCmp: {
2499 // Widen compares. Generate vector compares.
2500 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2501 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2502 setDebugLocFromInst(Builder, it);
2503 VectorParts &A = getVectorValue(it->getOperand(0));
2504 VectorParts &B = getVectorValue(it->getOperand(1));
2505 for (unsigned Part = 0; Part < UF; ++Part) {
2508 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2510 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2516 case Instruction::Store:
2517 case Instruction::Load:
2518 vectorizeMemoryInstruction(it, Legal);
2520 case Instruction::ZExt:
2521 case Instruction::SExt:
2522 case Instruction::FPToUI:
2523 case Instruction::FPToSI:
2524 case Instruction::FPExt:
2525 case Instruction::PtrToInt:
2526 case Instruction::IntToPtr:
2527 case Instruction::SIToFP:
2528 case Instruction::UIToFP:
2529 case Instruction::Trunc:
2530 case Instruction::FPTrunc:
2531 case Instruction::BitCast: {
2532 CastInst *CI = dyn_cast<CastInst>(it);
2533 setDebugLocFromInst(Builder, it);
2534 /// Optimize the special case where the source is the induction
2535 /// variable. Notice that we can only optimize the 'trunc' case
2536 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2537 /// c. other casts depend on pointer size.
2538 if (CI->getOperand(0) == OldInduction &&
2539 it->getOpcode() == Instruction::Trunc) {
2540 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2542 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2543 for (unsigned Part = 0; Part < UF; ++Part)
2544 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2547 /// Vectorize casts.
2548 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2550 VectorParts &A = getVectorValue(it->getOperand(0));
2551 for (unsigned Part = 0; Part < UF; ++Part)
2552 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2556 case Instruction::Call: {
2557 // Ignore dbg intrinsics.
2558 if (isa<DbgInfoIntrinsic>(it))
2560 setDebugLocFromInst(Builder, it);
2562 Module *M = BB->getParent()->getParent();
2563 CallInst *CI = cast<CallInst>(it);
2564 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2565 assert(ID && "Not an intrinsic call!");
2566 for (unsigned Part = 0; Part < UF; ++Part) {
2567 SmallVector<Value*, 4> Args;
2568 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2569 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2570 Args.push_back(Arg[Part]);
2572 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2573 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2574 Entry[Part] = Builder.CreateCall(F, Args);
2580 // All other instructions are unsupported. Scalarize them.
2581 scalarizeInstruction(it);
2584 }// end of for_each instr.
2587 void InnerLoopVectorizer::updateAnalysis() {
2588 // Forget the original basic block.
2589 SE->forgetLoop(OrigLoop);
2591 // Update the dominator tree information.
2592 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2593 "Entry does not dominate exit.");
2595 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2596 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2597 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2598 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2599 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2600 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2601 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2602 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2604 DEBUG(DT->verifyAnalysis());
2607 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2608 if (!EnableIfConversion)
2611 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2612 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2614 // Collect the blocks that need predication.
2615 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2616 BasicBlock *BB = LoopBlocks[i];
2618 // We don't support switch statements inside loops.
2619 if (!isa<BranchInst>(BB->getTerminator()))
2622 // We must be able to predicate all blocks that need to be predicated.
2623 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2627 // Check that we can actually speculate the hoistable loads.
2628 if (!LoadSpeculation.canHoistAllLoads())
2631 // We can if-convert this loop.
2635 bool LoopVectorizationLegality::canVectorize() {
2636 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2637 // be canonicalized.
2638 if (!TheLoop->getLoopPreheader())
2641 // We can only vectorize innermost loops.
2642 if (TheLoop->getSubLoopsVector().size())
2645 // We must have a single backedge.
2646 if (TheLoop->getNumBackEdges() != 1)
2649 // We must have a single exiting block.
2650 if (!TheLoop->getExitingBlock())
2653 unsigned NumBlocks = TheLoop->getNumBlocks();
2655 // Check if we can if-convert non single-bb loops.
2656 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2657 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2661 // We need to have a loop header.
2662 BasicBlock *Latch = TheLoop->getLoopLatch();
2663 DEBUG(dbgs() << "LV: Found a loop: " <<
2664 TheLoop->getHeader()->getName() << "\n");
2666 // ScalarEvolution needs to be able to find the exit count.
2667 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2668 if (ExitCount == SE->getCouldNotCompute()) {
2669 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2673 // Do not loop-vectorize loops with a tiny trip count.
2674 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2675 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2676 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2677 "This loop is not worth vectorizing.\n");
2681 // Check if we can vectorize the instructions and CFG in this loop.
2682 if (!canVectorizeInstrs()) {
2683 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2687 // Go over each instruction and look at memory deps.
2688 if (!canVectorizeMemory()) {
2689 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2693 // Collect all of the variables that remain uniform after vectorization.
2694 collectLoopUniforms();
2696 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2697 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2700 // Okay! We can vectorize. At this point we don't have any other mem analysis
2701 // which may limit our maximum vectorization factor, so just return true with
2706 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2707 if (Ty->isPointerTy())
2708 return DL.getIntPtrType(Ty->getContext());
2712 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2713 Ty0 = convertPointerToIntegerType(DL, Ty0);
2714 Ty1 = convertPointerToIntegerType(DL, Ty1);
2715 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2720 /// \brief Check that the instruction has outside loop users and is not an
2721 /// identified reduction variable.
2722 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2723 SmallPtrSet<Value *, 4> &Reductions) {
2724 // Reduction instructions are allowed to have exit users. All other
2725 // instructions must not have external users.
2726 if (!Reductions.count(Inst))
2727 //Check that all of the users of the loop are inside the BB.
2728 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2730 Instruction *U = cast<Instruction>(*I);
2731 // This user may be a reduction exit value.
2732 if (!TheLoop->contains(U)) {
2733 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2740 bool LoopVectorizationLegality::canVectorizeInstrs() {
2741 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2742 BasicBlock *Header = TheLoop->getHeader();
2744 // Look for the attribute signaling the absence of NaNs.
2745 Function &F = *Header->getParent();
2746 if (F.hasFnAttribute("no-nans-fp-math"))
2747 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2748 AttributeSet::FunctionIndex,
2749 "no-nans-fp-math").getValueAsString() == "true";
2751 // For each block in the loop.
2752 for (Loop::block_iterator bb = TheLoop->block_begin(),
2753 be = TheLoop->block_end(); bb != be; ++bb) {
2755 // Scan the instructions in the block and look for hazards.
2756 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2759 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2760 Type *PhiTy = Phi->getType();
2761 // Check that this PHI type is allowed.
2762 if (!PhiTy->isIntegerTy() &&
2763 !PhiTy->isFloatingPointTy() &&
2764 !PhiTy->isPointerTy()) {
2765 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2769 // If this PHINode is not in the header block, then we know that we
2770 // can convert it to select during if-conversion. No need to check if
2771 // the PHIs in this block are induction or reduction variables.
2772 if (*bb != Header) {
2773 // Check that this instruction has no outside users or is an
2774 // identified reduction value with an outside user.
2775 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2780 // We only allow if-converted PHIs with more than two incoming values.
2781 if (Phi->getNumIncomingValues() != 2) {
2782 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2786 // This is the value coming from the preheader.
2787 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2788 // Check if this is an induction variable.
2789 InductionKind IK = isInductionVariable(Phi);
2791 if (IK_NoInduction != IK) {
2792 // Get the widest type.
2794 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2796 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2798 // Int inductions are special because we only allow one IV.
2799 if (IK == IK_IntInduction) {
2800 // Use the phi node with the widest type as induction. Use the last
2801 // one if there are multiple (no good reason for doing this other
2802 // than it is expedient).
2803 if (!Induction || PhiTy == WidestIndTy)
2807 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2808 Inductions[Phi] = InductionInfo(StartValue, IK);
2812 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2813 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2816 if (AddReductionVar(Phi, RK_IntegerMult)) {
2817 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2820 if (AddReductionVar(Phi, RK_IntegerOr)) {
2821 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2824 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2825 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2828 if (AddReductionVar(Phi, RK_IntegerXor)) {
2829 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2832 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2833 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2836 if (AddReductionVar(Phi, RK_FloatMult)) {
2837 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2840 if (AddReductionVar(Phi, RK_FloatAdd)) {
2841 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2844 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2845 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2849 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2851 }// end of PHI handling
2853 // We still don't handle functions. However, we can ignore dbg intrinsic
2854 // calls and we do handle certain intrinsic and libm functions.
2855 CallInst *CI = dyn_cast<CallInst>(it);
2856 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2857 DEBUG(dbgs() << "LV: Found a call site.\n");
2861 // Check that the instruction return type is vectorizable.
2862 if (!VectorType::isValidElementType(it->getType()) &&
2863 !it->getType()->isVoidTy()) {
2864 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2868 // Check that the stored type is vectorizable.
2869 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2870 Type *T = ST->getValueOperand()->getType();
2871 if (!VectorType::isValidElementType(T))
2875 // Reduction instructions are allowed to have exit users.
2876 // All other instructions must not have external users.
2877 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2885 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2886 if (Inductions.empty())
2893 void LoopVectorizationLegality::collectLoopUniforms() {
2894 // We now know that the loop is vectorizable!
2895 // Collect variables that will remain uniform after vectorization.
2896 std::vector<Value*> Worklist;
2897 BasicBlock *Latch = TheLoop->getLoopLatch();
2899 // Start with the conditional branch and walk up the block.
2900 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2902 while (Worklist.size()) {
2903 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2904 Worklist.pop_back();
2906 // Look at instructions inside this loop.
2907 // Stop when reaching PHI nodes.
2908 // TODO: we need to follow values all over the loop, not only in this block.
2909 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2912 // This is a known uniform.
2915 // Insert all operands.
2916 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2920 /// \brief Analyses memory accesses in a loop.
2922 /// Checks whether run time pointer checks are needed and builds sets for data
2923 /// dependence checking.
2924 class AccessAnalysis {
2926 /// \brief Read or write access location.
2927 typedef std::pair<Value*, char> MemAccessInfo;
2929 /// \brief Set of potential dependent memory accesses.
2930 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2932 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2933 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2934 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2936 /// \brief Register a load and whether it is only read from.
2937 void addLoad(Value *Ptr, bool IsReadOnly) {
2938 Accesses.insert(std::make_pair(Ptr, false));
2940 ReadOnlyPtr.insert(Ptr);
2943 /// \brief Register a store.
2944 void addStore(Value *Ptr) {
2945 Accesses.insert(std::make_pair(Ptr, true));
2948 /// \brief Check whether we can check the pointers at runtime for
2949 /// non-intersection.
2950 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2951 unsigned &NumComparisons, ScalarEvolution *SE,
2954 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2955 /// and builds sets of dependent accesses.
2956 void buildDependenceSets() {
2957 // Process read-write pointers first.
2958 processMemAccesses(false);
2959 // Next, process read pointers.
2960 processMemAccesses(true);
2963 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
2965 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
2967 DenseSet<MemAccessInfo> &getDependenciesToCheck() { return CheckDeps; }
2970 typedef SetVector<MemAccessInfo> PtrAccessSet;
2971 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
2973 /// \brief Go over all memory access or only the deferred ones if
2974 /// \p UseDeferred is true and check whether runtime pointer checks are needed
2975 /// and build sets of dependency check candidates.
2976 void processMemAccesses(bool UseDeferred);
2978 /// Set of all accesses.
2979 PtrAccessSet Accesses;
2981 /// Set of access to check after all writes have been processed.
2982 PtrAccessSet DeferredAccesses;
2984 /// Map of pointers to last access encountered.
2985 UnderlyingObjToAccessMap ObjToLastAccess;
2987 /// Set of accesses that need a further dependence check.
2988 DenseSet<MemAccessInfo> CheckDeps;
2990 /// Set of pointers that are read only.
2991 SmallPtrSet<Value*, 16> ReadOnlyPtr;
2993 /// Set of underlying objects already written to.
2994 SmallPtrSet<Value*, 16> WriteObjects;
2998 /// Sets of potentially dependent accesses - members of one set share an
2999 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3000 /// dependence check.
3001 DepCandidates &DepCands;
3003 bool AreAllWritesIdentified;
3004 bool AreAllReadsIdentified;
3005 bool IsRTCheckNeeded;
3008 /// \brief Check whether a pointer can participate in a runtime bounds check.
3009 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3010 const SCEV *PtrScev = SE->getSCEV(Ptr);
3011 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3015 return AR->isAffine();
3018 bool AccessAnalysis::canCheckPtrAtRT(
3019 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3020 unsigned &NumComparisons, ScalarEvolution *SE,
3022 // Find pointers with computable bounds. We are going to use this information
3023 // to place a runtime bound check.
3024 unsigned NumReadPtrChecks = 0;
3025 unsigned NumWritePtrChecks = 0;
3026 bool CanDoRT = true;
3028 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3029 // We assign consecutive id to access from different dependence sets.
3030 // Accesses within the same set don't need a runtime check.
3031 unsigned RunningDepId = 1;
3032 DenseMap<Value *, unsigned> DepSetId;
3034 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3036 const MemAccessInfo &Access = *AI;
3037 Value *Ptr = Access.first;
3038 bool IsWrite = Access.second;
3040 // Just add write checks if we have both.
3041 if (!IsWrite && Accesses.count(std::make_pair(Ptr, true)))
3045 ++NumWritePtrChecks;
3049 if (hasComputableBounds(SE, Ptr)) {
3050 // The id of the dependence set.
3053 if (IsDepCheckNeeded) {
3054 Value *Leader = DepCands.getLeaderValue(Access).first;
3055 unsigned &LeaderId = DepSetId[Leader];
3057 LeaderId = RunningDepId++;
3060 // Each access has its own dependence set.
3061 DepId = RunningDepId++;
3063 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3065 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3071 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3072 NumComparisons = 0; // Only one dependence set.
3074 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3075 NumWritePtrChecks - 1));
3079 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3080 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3083 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3084 // We process the set twice: first we process read-write pointers, last we
3085 // process read-only pointers. This allows us to skip dependence tests for
3086 // read-only pointers.
3088 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3089 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3090 const MemAccessInfo &Access = *AI;
3091 Value *Ptr = Access.first;
3092 bool IsWrite = Access.second;
3094 DepCands.insert(Access);
3096 // Memorize read-only pointers for later processing and skip them in the
3097 // first round (they need to be checked after we have seen all write
3098 // pointers). Note: we also mark pointer that are not consecutive as
3099 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3100 // second check for "!IsWrite".
3101 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3102 if (!UseDeferred && IsReadOnlyPtr) {
3103 DeferredAccesses.insert(Access);
3107 bool NeedDepCheck = false;
3108 // Check whether there is the possiblity of dependency because of underlying
3109 // objects being the same.
3110 typedef SmallVector<Value*, 16> ValueVector;
3111 ValueVector TempObjects;
3112 GetUnderlyingObjects(Ptr, TempObjects, DL);
3113 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3115 Value *UnderlyingObj = *UI;
3117 // If this is a write then it needs to be an identified object. If this a
3118 // read and all writes (so far) are identified function scope objects we
3119 // don't need an identified underlying object but only an Argument (the
3120 // next write is going to invalidate this assumption if it is
3122 // This is a micro-optimization for the case where all writes are
3123 // identified and we have one argument pointer.
3124 // Otherwise, we do need a runtime check.
3125 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3126 (!IsWrite && (!AreAllWritesIdentified ||
3127 !isa<Argument>(UnderlyingObj)) &&
3128 !isIdentifiedObject(UnderlyingObj))) {
3129 DEBUG(dbgs() << "LV: Found an unidentified " <<
3130 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3132 IsRTCheckNeeded = (IsRTCheckNeeded ||
3133 !isIdentifiedObject(UnderlyingObj) ||
3134 !AreAllReadsIdentified);
3137 AreAllWritesIdentified = false;
3139 AreAllReadsIdentified = false;
3142 // If this is a write - check other reads and writes for conflicts. If
3143 // this is a read only check other writes for conflicts (but only if there
3144 // is no other write to the ptr - this is an optimization to catch "a[i] =
3145 // a[i] + " without having to do a dependence check).
3146 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3147 NeedDepCheck = true;
3150 WriteObjects.insert(UnderlyingObj);
3152 // Create sets of pointers connected by shared underlying objects.
3153 UnderlyingObjToAccessMap::iterator Prev =
3154 ObjToLastAccess.find(UnderlyingObj);
3155 if (Prev != ObjToLastAccess.end())
3156 DepCands.unionSets(Access, Prev->second);
3158 ObjToLastAccess[UnderlyingObj] = Access;
3162 CheckDeps.insert(Access);
3166 /// \brief Checks memory dependences among accesses to the same underlying
3167 /// object to determine whether there vectorization is legal or not (and at
3168 /// which vectorization factor).
3170 /// This class works under the assumption that we already checked that memory
3171 /// locations with different underlying pointers are "must-not alias".
3172 /// We use the ScalarEvolution framework to symbolically evalutate access
3173 /// functions pairs. Since we currently don't restructure the loop we can rely
3174 /// on the program order of memory accesses to determine their safety.
3175 /// At the moment we will only deem accesses as safe for:
3176 /// * A negative constant distance assuming program order.
3178 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3179 /// a[i] = tmp; y = a[i];
3181 /// The latter case is safe because later checks guarantuee that there can't
3182 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3183 /// the same variable: a header phi can only be an induction or a reduction, a
3184 /// reduction can't have a memory sink, an induction can't have a memory
3185 /// source). This is important and must not be violated (or we have to
3186 /// resort to checking for cycles through memory).
3188 /// * A positive constant distance assuming program order that is bigger
3189 /// than the biggest memory access.
3191 /// tmp = a[i] OR b[i] = x
3192 /// a[i+2] = tmp y = b[i+2];
3194 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3196 /// * Zero distances and all accesses have the same size.
3198 class MemoryDepChecker {
3200 typedef std::pair<Value*, char> MemAccessInfo;
3202 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3203 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3205 /// \brief Register the location (instructions are given increasing numbers)
3206 /// of a write access.
3207 void addAccess(StoreInst *SI) {
3208 Value *Ptr = SI->getPointerOperand();
3209 Accesses[std::make_pair(Ptr, true)].push_back(AccessIdx);
3210 InstMap.push_back(SI);
3214 /// \brief Register the location (instructions are given increasing numbers)
3215 /// of a write access.
3216 void addAccess(LoadInst *LI) {
3217 Value *Ptr = LI->getPointerOperand();
3218 Accesses[std::make_pair(Ptr, false)].push_back(AccessIdx);
3219 InstMap.push_back(LI);
3223 /// \brief Check whether the dependencies between the accesses are safe.
3225 /// Only checks sets with elements in \p CheckDeps.
3226 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3227 DenseSet<MemAccessInfo> &CheckDeps);
3229 /// \brief The maximum number of bytes of a vector register we can vectorize
3230 /// the accesses safely with.
3231 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3234 ScalarEvolution *SE;
3236 const Loop *InnermostLoop;
3238 /// \brief Maps access locations (ptr, read/write) to program order.
3239 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3241 /// \brief Memory access instructions in program order.
3242 SmallVector<Instruction *, 16> InstMap;
3244 /// \brief The program order index to be used for the next instruction.
3247 // We can access this many bytes in parallel safely.
3248 unsigned MaxSafeDepDistBytes;
3250 /// \brief Check whether there is a plausible dependence between the two
3253 /// Access \p A must happen before \p B in program order. The two indices
3254 /// identify the index into the program order map.
3256 /// This function checks whether there is a plausible dependence (or the
3257 /// absence of such can't be proved) between the two accesses. If there is a
3258 /// plausible dependence but the dependence distance is bigger than one
3259 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3260 /// distance is smaller than any other distance encountered so far).
3261 /// Otherwise, this function returns true signaling a possible dependence.
3262 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3263 const MemAccessInfo &B, unsigned BIdx);
3265 /// \brief Check whether the data dependence could prevent store-load
3267 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3270 static bool isInBoundsGep(Value *Ptr) {
3271 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3272 return GEP->isInBounds();
3276 /// \brief Check whether the access through \p Ptr has a constant stride.
3277 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3279 const Type *PtrTy = Ptr->getType();
3280 assert(PtrTy->isPointerTy() && "Unexpected non ptr");
3282 // Make sure that the pointer does not point to aggregate types.
3283 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) {
3284 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr
3289 const SCEV *PtrScev = SE->getSCEV(Ptr);
3290 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3292 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3293 << *Ptr << " SCEV: " << *PtrScev << "\n");
3297 // The accesss function must stride over the innermost loop.
3298 if (Lp != AR->getLoop()) {
3299 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << *Ptr
3300 << " SCEV: " << *PtrScev << "\n");
3303 // The address calculation must not wrap. Otherwise, a dependence could be
3304 // inverted. An inbounds getelementptr that is a AddRec with a unit stride
3305 // cannot wrap per definition. The unit stride requirement is checked later.
3306 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3307 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3308 if (!IsNoWrapAddRec && !IsInBoundsGEP) {
3309 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3310 << *Ptr << " SCEV: " << *PtrScev << "\n");
3314 // Check the step is constant.
3315 const SCEV *Step = AR->getStepRecurrence(*SE);
3317 // Calculate the pointer stride and check if it is consecutive.
3318 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3320 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3321 " SCEV: " << *PtrScev << "\n");
3325 int64_t Size = DL->getTypeAllocSize(PtrTy->getPointerElementType());
3326 const APInt &APStepVal = C->getValue()->getValue();
3328 // Huge step value - give up.
3329 if (APStepVal.getBitWidth() > 64)
3332 int64_t StepVal = APStepVal.getSExtValue();
3335 int64_t Stride = StepVal / Size;
3336 int64_t Rem = StepVal % Size;
3340 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3341 // know we can't "wrap around the address space".
3342 if (!IsNoWrapAddRec && IsInBoundsGEP && Stride != 1 && Stride != -1)
3348 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3349 unsigned TypeByteSize) {
3350 // If loads occur at a distance that is not a multiple of a feasible vector
3351 // factor store-load forwarding does not take place.
3352 // Positive dependences might cause troubles because vectorizing them might
3353 // prevent store-load forwarding making vectorized code run a lot slower.
3354 // a[i] = a[i-3] ^ a[i-8];
3355 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3356 // hence on your typical architecture store-load forwarding does not take
3357 // place. Vectorizing in such cases does not make sense.
3358 // Store-load forwarding distance.
3359 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3360 // Maximum vector factor.
3361 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3362 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3363 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3365 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3367 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3368 MaxVFWithoutSLForwardIssues = (vf >>=1);
3373 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3374 DEBUG(dbgs() << "LV: Distance " << Distance <<
3375 " that could cause a store-load forwarding conflict\n");
3379 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3380 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3381 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3385 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3386 const MemAccessInfo &B, unsigned BIdx) {
3387 assert (AIdx < BIdx && "Must pass arguments in program order");
3389 Value *APtr = A.first;
3390 Value *BPtr = B.first;
3391 bool AIsWrite = A.second;
3392 bool BIsWrite = B.second;
3394 // Two reads are independent.
3395 if (!AIsWrite && !BIsWrite)
3398 const SCEV *AScev = SE->getSCEV(APtr);
3399 const SCEV *BScev = SE->getSCEV(BPtr);
3401 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3402 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3404 const SCEV *Src = AScev;
3405 const SCEV *Sink = BScev;
3407 // If the induction step is negative we have to invert source and sink of the
3409 if (StrideAPtr < 0) {
3412 std::swap(APtr, BPtr);
3413 std::swap(Src, Sink);
3414 std::swap(AIsWrite, BIsWrite);
3415 std::swap(AIdx, BIdx);
3416 std::swap(StrideAPtr, StrideBPtr);
3419 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3421 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3422 << "(Induction step: " << StrideAPtr << ")\n");
3423 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3424 << *InstMap[BIdx] << ": " << *Dist << "\n");
3426 // Need consecutive accesses. We don't want to vectorize
3427 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3428 // the address space.
3429 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3430 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3434 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3436 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3440 Type *ATy = APtr->getType()->getPointerElementType();
3441 Type *BTy = BPtr->getType()->getPointerElementType();
3442 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3444 // Negative distances are not plausible dependencies.
3445 const APInt &Val = C->getValue()->getValue();
3446 if (Val.isNegative()) {
3447 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3448 if (IsTrueDataDependence &&
3449 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3453 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3457 // Write to the same location with the same size.
3458 // Could be improved to assert type sizes are the same (i32 == float, etc).
3462 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3466 assert(Val.isStrictlyPositive() && "Expect a positive value");
3468 // Positive distance bigger than max vectorization factor.
3471 "LV: ReadWrite-Write positive dependency with different types");
3475 unsigned Distance = (unsigned) Val.getZExtValue();
3477 // Bail out early if passed-in parameters make vectorization not feasible.
3478 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3479 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3481 // The distance must be bigger than the size needed for a vectorized version
3482 // of the operation and the size of the vectorized operation must not be
3483 // bigger than the currrent maximum size.
3484 if (Distance < 2*TypeByteSize ||
3485 2*TypeByteSize > MaxSafeDepDistBytes ||
3486 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3487 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3488 << Val.getSExtValue() << "\n");
3492 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3493 Distance : MaxSafeDepDistBytes;
3495 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3496 if (IsTrueDataDependence &&
3497 couldPreventStoreLoadForward(Distance, TypeByteSize))
3500 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3501 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3507 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3508 DenseSet<MemAccessInfo> &CheckDeps) {
3510 MaxSafeDepDistBytes = -1U;
3511 while (!CheckDeps.empty()) {
3512 MemAccessInfo CurAccess = *CheckDeps.begin();
3514 // Get the relevant memory access set.
3515 EquivalenceClasses<MemAccessInfo>::iterator I =
3516 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3518 // Check accesses within this set.
3519 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3520 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3522 // Check every access pair.
3524 CheckDeps.erase(*AI);
3525 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3527 // Check every accessing instruction pair in program order.
3528 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3529 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3530 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3531 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3532 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3534 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3545 bool LoopVectorizationLegality::canVectorizeMemory() {
3547 typedef SmallVector<Value*, 16> ValueVector;
3548 typedef SmallPtrSet<Value*, 16> ValueSet;
3550 // Stores a pair of memory access location and whether the access is a store
3551 // (true) or a load (false).
3552 typedef std::pair<Value*, char> MemAccessInfo;
3553 typedef DenseSet<MemAccessInfo> PtrAccessSet;
3555 // Holds the Load and Store *instructions*.
3559 // Holds all the different accesses in the loop.
3560 unsigned NumReads = 0;
3561 unsigned NumReadWrites = 0;
3563 PtrRtCheck.Pointers.clear();
3564 PtrRtCheck.Need = false;
3566 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3567 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3570 for (Loop::block_iterator bb = TheLoop->block_begin(),
3571 be = TheLoop->block_end(); bb != be; ++bb) {
3573 // Scan the BB and collect legal loads and stores.
3574 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3577 // If this is a load, save it. If this instruction can read from memory
3578 // but is not a load, then we quit. Notice that we don't handle function
3579 // calls that read or write.
3580 if (it->mayReadFromMemory()) {
3581 LoadInst *Ld = dyn_cast<LoadInst>(it);
3582 if (!Ld) return false;
3583 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3584 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3587 Loads.push_back(Ld);
3588 DepChecker.addAccess(Ld);
3592 // Save 'store' instructions. Abort if other instructions write to memory.
3593 if (it->mayWriteToMemory()) {
3594 StoreInst *St = dyn_cast<StoreInst>(it);
3595 if (!St) return false;
3596 if (!St->isSimple() && !IsAnnotatedParallel) {
3597 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3600 Stores.push_back(St);
3601 DepChecker.addAccess(St);
3606 // Now we have two lists that hold the loads and the stores.
3607 // Next, we find the pointers that they use.
3609 // Check if we see any stores. If there are no stores, then we don't
3610 // care if the pointers are *restrict*.
3611 if (!Stores.size()) {
3612 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3616 AccessAnalysis::DepCandidates DependentAccesses;
3617 AccessAnalysis Accesses(DL, DependentAccesses);
3619 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3620 // multiple times on the same object. If the ptr is accessed twice, once
3621 // for read and once for write, it will only appear once (on the write
3622 // list). This is okay, since we are going to check for conflicts between
3623 // writes and between reads and writes, but not between reads and reads.
3626 ValueVector::iterator I, IE;
3627 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3628 StoreInst *ST = cast<StoreInst>(*I);
3629 Value* Ptr = ST->getPointerOperand();
3631 if (isUniform(Ptr)) {
3632 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3636 // If we did *not* see this pointer before, insert it to the read-write
3637 // list. At this phase it is only a 'write' list.
3638 if (Seen.insert(Ptr)) {
3640 Accesses.addStore(Ptr);
3644 if (IsAnnotatedParallel) {
3646 << "LV: A loop annotated parallel, ignore memory dependency "
3651 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3652 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3653 LoadInst *LD = cast<LoadInst>(*I);
3654 Value* Ptr = LD->getPointerOperand();
3655 // If we did *not* see this pointer before, insert it to the
3656 // read list. If we *did* see it before, then it is already in
3657 // the read-write list. This allows us to vectorize expressions
3658 // such as A[i] += x; Because the address of A[i] is a read-write
3659 // pointer. This only works if the index of A[i] is consecutive.
3660 // If the address of i is unknown (for example A[B[i]]) then we may
3661 // read a few words, modify, and write a few words, and some of the
3662 // words may be written to the same address.
3663 bool IsReadOnlyPtr = false;
3664 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3666 IsReadOnlyPtr = true;
3668 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3671 // If we write (or read-write) to a single destination and there are no
3672 // other reads in this loop then is it safe to vectorize.
3673 if (NumReadWrites == 1 && NumReads == 0) {
3674 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3678 // Build dependence sets and check whether we need a runtime pointer bounds
3680 Accesses.buildDependenceSets();
3681 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3683 // Find pointers with computable bounds. We are going to use this information
3684 // to place a runtime bound check.
3685 unsigned NumComparisons = 0;
3686 bool CanDoRT = false;
3688 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3691 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3692 " pointer comparisons.\n");
3694 // If we only have one set of dependences to check pointers among we don't
3695 // need a runtime check.
3696 if (NumComparisons == 0 && NeedRTCheck)
3697 NeedRTCheck = false;
3699 // Check that we did not collect too many pointers or found a unsizeable
3701 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3707 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3710 if (NeedRTCheck && !CanDoRT) {
3711 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3712 "the array bounds.\n");
3717 PtrRtCheck.Need = NeedRTCheck;
3719 bool CanVecMem = true;
3720 if (Accesses.isDependencyCheckNeeded()) {
3721 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3722 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3723 Accesses.getDependenciesToCheck());
3724 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3727 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3728 " need a runtime memory check.\n");
3733 static bool hasMultipleUsesOf(Instruction *I,
3734 SmallPtrSet<Instruction *, 8> &Insts) {
3735 unsigned NumUses = 0;
3736 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3737 if (Insts.count(dyn_cast<Instruction>(*Use)))
3746 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3747 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3748 if (!Set.count(dyn_cast<Instruction>(*Use)))
3753 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3754 ReductionKind Kind) {
3755 if (Phi->getNumIncomingValues() != 2)
3758 // Reduction variables are only found in the loop header block.
3759 if (Phi->getParent() != TheLoop->getHeader())
3762 // Obtain the reduction start value from the value that comes from the loop
3764 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3766 // ExitInstruction is the single value which is used outside the loop.
3767 // We only allow for a single reduction value to be used outside the loop.
3768 // This includes users of the reduction, variables (which form a cycle
3769 // which ends in the phi node).
3770 Instruction *ExitInstruction = 0;
3771 // Indicates that we found a reduction operation in our scan.
3772 bool FoundReduxOp = false;
3774 // We start with the PHI node and scan for all of the users of this
3775 // instruction. All users must be instructions that can be used as reduction
3776 // variables (such as ADD). We must have a single out-of-block user. The cycle
3777 // must include the original PHI.
3778 bool FoundStartPHI = false;
3780 // To recognize min/max patterns formed by a icmp select sequence, we store
3781 // the number of instruction we saw from the recognized min/max pattern,
3782 // to make sure we only see exactly the two instructions.
3783 unsigned NumCmpSelectPatternInst = 0;
3784 ReductionInstDesc ReduxDesc(false, 0);
3786 SmallPtrSet<Instruction *, 8> VisitedInsts;
3787 SmallVector<Instruction *, 8> Worklist;
3788 Worklist.push_back(Phi);
3789 VisitedInsts.insert(Phi);
3791 // A value in the reduction can be used:
3792 // - By the reduction:
3793 // - Reduction operation:
3794 // - One use of reduction value (safe).
3795 // - Multiple use of reduction value (not safe).
3797 // - All uses of the PHI must be the reduction (safe).
3798 // - Otherwise, not safe.
3799 // - By one instruction outside of the loop (safe).
3800 // - By further instructions outside of the loop (not safe).
3801 // - By an instruction that is not part of the reduction (not safe).
3803 // * An instruction type other than PHI or the reduction operation.
3804 // * A PHI in the header other than the initial PHI.
3805 while (!Worklist.empty()) {
3806 Instruction *Cur = Worklist.back();
3807 Worklist.pop_back();
3810 // If the instruction has no users then this is a broken chain and can't be
3811 // a reduction variable.
3812 if (Cur->use_empty())
3815 bool IsAPhi = isa<PHINode>(Cur);
3817 // A header PHI use other than the original PHI.
3818 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3821 // Reductions of instructions such as Div, and Sub is only possible if the
3822 // LHS is the reduction variable.
3823 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3824 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3825 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3828 // Any reduction instruction must be of one of the allowed kinds.
3829 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3830 if (!ReduxDesc.IsReduction)
3833 // A reduction operation must only have one use of the reduction value.
3834 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3835 hasMultipleUsesOf(Cur, VisitedInsts))
3838 // All inputs to a PHI node must be a reduction value.
3839 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3842 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3843 isa<SelectInst>(Cur)))
3844 ++NumCmpSelectPatternInst;
3845 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3846 isa<SelectInst>(Cur)))
3847 ++NumCmpSelectPatternInst;
3849 // Check whether we found a reduction operator.
3850 FoundReduxOp |= !IsAPhi;
3852 // Process users of current instruction. Push non PHI nodes after PHI nodes
3853 // onto the stack. This way we are going to have seen all inputs to PHI
3854 // nodes once we get to them.
3855 SmallVector<Instruction *, 8> NonPHIs;
3856 SmallVector<Instruction *, 8> PHIs;
3857 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3859 Instruction *Usr = cast<Instruction>(*UI);
3861 // Check if we found the exit user.
3862 BasicBlock *Parent = Usr->getParent();
3863 if (!TheLoop->contains(Parent)) {
3864 // Exit if you find multiple outside users.
3865 if (ExitInstruction != 0)
3867 ExitInstruction = Cur;
3871 // Process instructions only once (termination).
3872 if (VisitedInsts.insert(Usr)) {
3873 if (isa<PHINode>(Usr))
3874 PHIs.push_back(Usr);
3876 NonPHIs.push_back(Usr);
3878 // Remember that we completed the cycle.
3880 FoundStartPHI = true;
3882 Worklist.append(PHIs.begin(), PHIs.end());
3883 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3886 // This means we have seen one but not the other instruction of the
3887 // pattern or more than just a select and cmp.
3888 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3889 NumCmpSelectPatternInst != 2)
3892 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3895 // We found a reduction var if we have reached the original phi node and we
3896 // only have a single instruction with out-of-loop users.
3898 // This instruction is allowed to have out-of-loop users.
3899 AllowedExit.insert(ExitInstruction);
3901 // Save the description of this reduction variable.
3902 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3903 ReduxDesc.MinMaxKind);
3904 Reductions[Phi] = RD;
3905 // We've ended the cycle. This is a reduction variable if we have an
3906 // outside user and it has a binary op.
3911 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3912 /// pattern corresponding to a min(X, Y) or max(X, Y).
3913 LoopVectorizationLegality::ReductionInstDesc
3914 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3915 ReductionInstDesc &Prev) {
3917 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3918 "Expect a select instruction");
3919 Instruction *Cmp = 0;
3920 SelectInst *Select = 0;
3922 // We must handle the select(cmp()) as a single instruction. Advance to the
3924 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3925 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3926 return ReductionInstDesc(false, I);
3927 return ReductionInstDesc(Select, Prev.MinMaxKind);
3930 // Only handle single use cases for now.
3931 if (!(Select = dyn_cast<SelectInst>(I)))
3932 return ReductionInstDesc(false, I);
3933 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3934 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3935 return ReductionInstDesc(false, I);
3936 if (!Cmp->hasOneUse())
3937 return ReductionInstDesc(false, I);
3942 // Look for a min/max pattern.
3943 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3944 return ReductionInstDesc(Select, MRK_UIntMin);
3945 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3946 return ReductionInstDesc(Select, MRK_UIntMax);
3947 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3948 return ReductionInstDesc(Select, MRK_SIntMax);
3949 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3950 return ReductionInstDesc(Select, MRK_SIntMin);
3951 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3952 return ReductionInstDesc(Select, MRK_FloatMin);
3953 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3954 return ReductionInstDesc(Select, MRK_FloatMax);
3955 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3956 return ReductionInstDesc(Select, MRK_FloatMin);
3957 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3958 return ReductionInstDesc(Select, MRK_FloatMax);
3960 return ReductionInstDesc(false, I);
3963 LoopVectorizationLegality::ReductionInstDesc
3964 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3966 ReductionInstDesc &Prev) {
3967 bool FP = I->getType()->isFloatingPointTy();
3968 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3969 switch (I->getOpcode()) {
3971 return ReductionInstDesc(false, I);
3972 case Instruction::PHI:
3973 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3974 Kind != RK_FloatMinMax))
3975 return ReductionInstDesc(false, I);
3976 return ReductionInstDesc(I, Prev.MinMaxKind);
3977 case Instruction::Sub:
3978 case Instruction::Add:
3979 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3980 case Instruction::Mul:
3981 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3982 case Instruction::And:
3983 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3984 case Instruction::Or:
3985 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3986 case Instruction::Xor:
3987 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3988 case Instruction::FMul:
3989 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3990 case Instruction::FAdd:
3991 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3992 case Instruction::FCmp:
3993 case Instruction::ICmp:
3994 case Instruction::Select:
3995 if (Kind != RK_IntegerMinMax &&
3996 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3997 return ReductionInstDesc(false, I);
3998 return isMinMaxSelectCmpPattern(I, Prev);
4002 LoopVectorizationLegality::InductionKind
4003 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4004 Type *PhiTy = Phi->getType();
4005 // We only handle integer and pointer inductions variables.
4006 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4007 return IK_NoInduction;
4009 // Check that the PHI is consecutive.
4010 const SCEV *PhiScev = SE->getSCEV(Phi);
4011 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4013 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4014 return IK_NoInduction;
4016 const SCEV *Step = AR->getStepRecurrence(*SE);
4018 // Integer inductions need to have a stride of one.
4019 if (PhiTy->isIntegerTy()) {
4021 return IK_IntInduction;
4022 if (Step->isAllOnesValue())
4023 return IK_ReverseIntInduction;
4024 return IK_NoInduction;
4027 // Calculate the pointer stride and check if it is consecutive.
4028 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4030 return IK_NoInduction;
4032 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4033 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4034 if (C->getValue()->equalsInt(Size))
4035 return IK_PtrInduction;
4036 else if (C->getValue()->equalsInt(0 - Size))
4037 return IK_ReversePtrInduction;
4039 return IK_NoInduction;
4042 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4043 Value *In0 = const_cast<Value*>(V);
4044 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4048 return Inductions.count(PN);
4051 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4052 assert(TheLoop->contains(BB) && "Unknown block used");
4054 // Blocks that do not dominate the latch need predication.
4055 BasicBlock* Latch = TheLoop->getLoopLatch();
4056 return !DT->dominates(BB, Latch);
4059 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
4060 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4061 // We might be able to hoist the load.
4062 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
4065 // We don't predicate stores at the moment.
4066 if (it->mayWriteToMemory() || it->mayThrow())
4069 // The instructions below can trap.
4070 switch (it->getOpcode()) {
4072 case Instruction::UDiv:
4073 case Instruction::SDiv:
4074 case Instruction::URem:
4075 case Instruction::SRem:
4083 LoopVectorizationCostModel::VectorizationFactor
4084 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4086 // Width 1 means no vectorize
4087 VectorizationFactor Factor = { 1U, 0U };
4088 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4089 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4093 // Find the trip count.
4094 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4095 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4097 unsigned WidestType = getWidestType();
4098 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4099 unsigned MaxSafeDepDist = -1U;
4100 if (Legal->getMaxSafeDepDistBytes() != -1U)
4101 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4102 WidestRegister = WidestRegister < MaxSafeDepDist ? WidestRegister : MaxSafeDepDist;
4103 unsigned MaxVectorSize = WidestRegister / WidestType;
4104 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4105 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4107 if (MaxVectorSize == 0) {
4108 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4112 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4113 " into one vector!");
4115 unsigned VF = MaxVectorSize;
4117 // If we optimize the program for size, avoid creating the tail loop.
4119 // If we are unable to calculate the trip count then don't try to vectorize.
4121 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4125 // Find the maximum SIMD width that can fit within the trip count.
4126 VF = TC % MaxVectorSize;
4131 // If the trip count that we found modulo the vectorization factor is not
4132 // zero then we require a tail.
4134 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4140 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4141 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4143 Factor.Width = UserVF;
4147 float Cost = expectedCost(1);
4149 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4150 for (unsigned i=2; i <= VF; i*=2) {
4151 // Notice that the vector loop needs to be executed less times, so
4152 // we need to divide the cost of the vector loops by the width of
4153 // the vector elements.
4154 float VectorCost = expectedCost(i) / (float)i;
4155 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4156 (int)VectorCost << ".\n");
4157 if (VectorCost < Cost) {
4163 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4164 Factor.Width = Width;
4165 Factor.Cost = Width * Cost;
4169 unsigned LoopVectorizationCostModel::getWidestType() {
4170 unsigned MaxWidth = 8;
4173 for (Loop::block_iterator bb = TheLoop->block_begin(),
4174 be = TheLoop->block_end(); bb != be; ++bb) {
4175 BasicBlock *BB = *bb;
4177 // For each instruction in the loop.
4178 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4179 Type *T = it->getType();
4181 // Only examine Loads, Stores and PHINodes.
4182 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4185 // Examine PHI nodes that are reduction variables.
4186 if (PHINode *PN = dyn_cast<PHINode>(it))
4187 if (!Legal->getReductionVars()->count(PN))
4190 // Examine the stored values.
4191 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4192 T = ST->getValueOperand()->getType();
4194 // Ignore loaded pointer types and stored pointer types that are not
4195 // consecutive. However, we do want to take consecutive stores/loads of
4196 // pointer vectors into account.
4197 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4200 MaxWidth = std::max(MaxWidth,
4201 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4209 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4212 unsigned LoopCost) {
4214 // -- The unroll heuristics --
4215 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4216 // There are many micro-architectural considerations that we can't predict
4217 // at this level. For example frontend pressure (on decode or fetch) due to
4218 // code size, or the number and capabilities of the execution ports.
4220 // We use the following heuristics to select the unroll factor:
4221 // 1. If the code has reductions the we unroll in order to break the cross
4222 // iteration dependency.
4223 // 2. If the loop is really small then we unroll in order to reduce the loop
4225 // 3. We don't unroll if we think that we will spill registers to memory due
4226 // to the increased register pressure.
4228 // Use the user preference, unless 'auto' is selected.
4232 // When we optimize for size we don't unroll.
4236 // We used the distance for the unroll factor.
4237 if (Legal->getMaxSafeDepDistBytes() != -1U)
4240 // Do not unroll loops with a relatively small trip count.
4241 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4242 TheLoop->getLoopLatch());
4243 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4246 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4247 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4248 " vector registers\n");
4250 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4251 // We divide by these constants so assume that we have at least one
4252 // instruction that uses at least one register.
4253 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4254 R.NumInstructions = std::max(R.NumInstructions, 1U);
4256 // We calculate the unroll factor using the following formula.
4257 // Subtract the number of loop invariants from the number of available
4258 // registers. These registers are used by all of the unrolled instances.
4259 // Next, divide the remaining registers by the number of registers that is
4260 // required by the loop, in order to estimate how many parallel instances
4261 // fit without causing spills.
4262 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4264 // Clamp the unroll factor ranges to reasonable factors.
4265 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4267 // If we did not calculate the cost for VF (because the user selected the VF)
4268 // then we calculate the cost of VF here.
4270 LoopCost = expectedCost(VF);
4272 // Clamp the calculated UF to be between the 1 and the max unroll factor
4273 // that the target allows.
4274 if (UF > MaxUnrollSize)
4279 if (Legal->getReductionVars()->size()) {
4280 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4284 // We want to unroll tiny loops in order to reduce the loop overhead.
4285 // We assume that the cost overhead is 1 and we use the cost model
4286 // to estimate the cost of the loop and unroll until the cost of the
4287 // loop overhead is about 5% of the cost of the loop.
4288 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4289 if (LoopCost < 20) {
4290 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4291 unsigned NewUF = 20/LoopCost + 1;
4292 return std::min(NewUF, UF);
4295 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4299 LoopVectorizationCostModel::RegisterUsage
4300 LoopVectorizationCostModel::calculateRegisterUsage() {
4301 // This function calculates the register usage by measuring the highest number
4302 // of values that are alive at a single location. Obviously, this is a very
4303 // rough estimation. We scan the loop in a topological order in order and
4304 // assign a number to each instruction. We use RPO to ensure that defs are
4305 // met before their users. We assume that each instruction that has in-loop
4306 // users starts an interval. We record every time that an in-loop value is
4307 // used, so we have a list of the first and last occurrences of each
4308 // instruction. Next, we transpose this data structure into a multi map that
4309 // holds the list of intervals that *end* at a specific location. This multi
4310 // map allows us to perform a linear search. We scan the instructions linearly
4311 // and record each time that a new interval starts, by placing it in a set.
4312 // If we find this value in the multi-map then we remove it from the set.
4313 // The max register usage is the maximum size of the set.
4314 // We also search for instructions that are defined outside the loop, but are
4315 // used inside the loop. We need this number separately from the max-interval
4316 // usage number because when we unroll, loop-invariant values do not take
4318 LoopBlocksDFS DFS(TheLoop);
4322 R.NumInstructions = 0;
4324 // Each 'key' in the map opens a new interval. The values
4325 // of the map are the index of the 'last seen' usage of the
4326 // instruction that is the key.
4327 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4328 // Maps instruction to its index.
4329 DenseMap<unsigned, Instruction*> IdxToInstr;
4330 // Marks the end of each interval.
4331 IntervalMap EndPoint;
4332 // Saves the list of instruction indices that are used in the loop.
4333 SmallSet<Instruction*, 8> Ends;
4334 // Saves the list of values that are used in the loop but are
4335 // defined outside the loop, such as arguments and constants.
4336 SmallPtrSet<Value*, 8> LoopInvariants;
4339 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4340 be = DFS.endRPO(); bb != be; ++bb) {
4341 R.NumInstructions += (*bb)->size();
4342 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4344 Instruction *I = it;
4345 IdxToInstr[Index++] = I;
4347 // Save the end location of each USE.
4348 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4349 Value *U = I->getOperand(i);
4350 Instruction *Instr = dyn_cast<Instruction>(U);
4352 // Ignore non-instruction values such as arguments, constants, etc.
4353 if (!Instr) continue;
4355 // If this instruction is outside the loop then record it and continue.
4356 if (!TheLoop->contains(Instr)) {
4357 LoopInvariants.insert(Instr);
4361 // Overwrite previous end points.
4362 EndPoint[Instr] = Index;
4368 // Saves the list of intervals that end with the index in 'key'.
4369 typedef SmallVector<Instruction*, 2> InstrList;
4370 DenseMap<unsigned, InstrList> TransposeEnds;
4372 // Transpose the EndPoints to a list of values that end at each index.
4373 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4375 TransposeEnds[it->second].push_back(it->first);
4377 SmallSet<Instruction*, 8> OpenIntervals;
4378 unsigned MaxUsage = 0;
4381 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4382 for (unsigned int i = 0; i < Index; ++i) {
4383 Instruction *I = IdxToInstr[i];
4384 // Ignore instructions that are never used within the loop.
4385 if (!Ends.count(I)) continue;
4387 // Remove all of the instructions that end at this location.
4388 InstrList &List = TransposeEnds[i];
4389 for (unsigned int j=0, e = List.size(); j < e; ++j)
4390 OpenIntervals.erase(List[j]);
4392 // Count the number of live interals.
4393 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4395 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4396 OpenIntervals.size() <<"\n");
4398 // Add the current instruction to the list of open intervals.
4399 OpenIntervals.insert(I);
4402 unsigned Invariant = LoopInvariants.size();
4403 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4404 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4405 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4407 R.LoopInvariantRegs = Invariant;
4408 R.MaxLocalUsers = MaxUsage;
4412 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4416 for (Loop::block_iterator bb = TheLoop->block_begin(),
4417 be = TheLoop->block_end(); bb != be; ++bb) {
4418 unsigned BlockCost = 0;
4419 BasicBlock *BB = *bb;
4421 // For each instruction in the old loop.
4422 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4423 // Skip dbg intrinsics.
4424 if (isa<DbgInfoIntrinsic>(it))
4427 unsigned C = getInstructionCost(it, VF);
4429 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4430 VF << " For instruction: "<< *it << "\n");
4433 // We assume that if-converted blocks have a 50% chance of being executed.
4434 // When the code is scalar then some of the blocks are avoided due to CF.
4435 // When the code is vectorized we execute all code paths.
4436 if (Legal->blockNeedsPredication(*bb) && VF == 1)
4446 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4447 // If we know that this instruction will remain uniform, check the cost of
4448 // the scalar version.
4449 if (Legal->isUniformAfterVectorization(I))
4452 Type *RetTy = I->getType();
4453 Type *VectorTy = ToVectorTy(RetTy, VF);
4455 // TODO: We need to estimate the cost of intrinsic calls.
4456 switch (I->getOpcode()) {
4457 case Instruction::GetElementPtr:
4458 // We mark this instruction as zero-cost because the cost of GEPs in
4459 // vectorized code depends on whether the corresponding memory instruction
4460 // is scalarized or not. Therefore, we handle GEPs with the memory
4461 // instruction cost.
4463 case Instruction::Br: {
4464 return TTI.getCFInstrCost(I->getOpcode());
4466 case Instruction::PHI:
4467 //TODO: IF-converted IFs become selects.
4469 case Instruction::Add:
4470 case Instruction::FAdd:
4471 case Instruction::Sub:
4472 case Instruction::FSub:
4473 case Instruction::Mul:
4474 case Instruction::FMul:
4475 case Instruction::UDiv:
4476 case Instruction::SDiv:
4477 case Instruction::FDiv:
4478 case Instruction::URem:
4479 case Instruction::SRem:
4480 case Instruction::FRem:
4481 case Instruction::Shl:
4482 case Instruction::LShr:
4483 case Instruction::AShr:
4484 case Instruction::And:
4485 case Instruction::Or:
4486 case Instruction::Xor: {
4487 // Certain instructions can be cheaper to vectorize if they have a constant
4488 // second vector operand. One example of this are shifts on x86.
4489 TargetTransformInfo::OperandValueKind Op1VK =
4490 TargetTransformInfo::OK_AnyValue;
4491 TargetTransformInfo::OperandValueKind Op2VK =
4492 TargetTransformInfo::OK_AnyValue;
4494 if (isa<ConstantInt>(I->getOperand(1)))
4495 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4497 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4499 case Instruction::Select: {
4500 SelectInst *SI = cast<SelectInst>(I);
4501 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4502 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4503 Type *CondTy = SI->getCondition()->getType();
4505 CondTy = VectorType::get(CondTy, VF);
4507 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4509 case Instruction::ICmp:
4510 case Instruction::FCmp: {
4511 Type *ValTy = I->getOperand(0)->getType();
4512 VectorTy = ToVectorTy(ValTy, VF);
4513 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4515 case Instruction::Store:
4516 case Instruction::Load: {
4517 StoreInst *SI = dyn_cast<StoreInst>(I);
4518 LoadInst *LI = dyn_cast<LoadInst>(I);
4519 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4521 VectorTy = ToVectorTy(ValTy, VF);
4523 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4524 unsigned AS = SI ? SI->getPointerAddressSpace() :
4525 LI->getPointerAddressSpace();
4526 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4527 // We add the cost of address computation here instead of with the gep
4528 // instruction because only here we know whether the operation is
4531 return TTI.getAddressComputationCost(VectorTy) +
4532 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4534 // Scalarized loads/stores.
4535 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4536 bool Reverse = ConsecutiveStride < 0;
4537 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4538 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4539 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4541 // The cost of extracting from the value vector and pointer vector.
4542 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4543 for (unsigned i = 0; i < VF; ++i) {
4544 // The cost of extracting the pointer operand.
4545 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4546 // In case of STORE, the cost of ExtractElement from the vector.
4547 // In case of LOAD, the cost of InsertElement into the returned
4549 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4550 Instruction::InsertElement,
4554 // The cost of the scalar loads/stores.
4555 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
4556 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4561 // Wide load/stores.
4562 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4563 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4566 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4570 case Instruction::ZExt:
4571 case Instruction::SExt:
4572 case Instruction::FPToUI:
4573 case Instruction::FPToSI:
4574 case Instruction::FPExt:
4575 case Instruction::PtrToInt:
4576 case Instruction::IntToPtr:
4577 case Instruction::SIToFP:
4578 case Instruction::UIToFP:
4579 case Instruction::Trunc:
4580 case Instruction::FPTrunc:
4581 case Instruction::BitCast: {
4582 // We optimize the truncation of induction variable.
4583 // The cost of these is the same as the scalar operation.
4584 if (I->getOpcode() == Instruction::Trunc &&
4585 Legal->isInductionVariable(I->getOperand(0)))
4586 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4587 I->getOperand(0)->getType());
4589 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4590 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4592 case Instruction::Call: {
4593 CallInst *CI = cast<CallInst>(I);
4594 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4595 assert(ID && "Not an intrinsic call!");
4596 Type *RetTy = ToVectorTy(CI->getType(), VF);
4597 SmallVector<Type*, 4> Tys;
4598 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4599 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4600 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4603 // We are scalarizing the instruction. Return the cost of the scalar
4604 // instruction, plus the cost of insert and extract into vector
4605 // elements, times the vector width.
4608 if (!RetTy->isVoidTy() && VF != 1) {
4609 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4611 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4614 // The cost of inserting the results plus extracting each one of the
4616 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4619 // The cost of executing VF copies of the scalar instruction. This opcode
4620 // is unknown. Assume that it is the same as 'mul'.
4621 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4627 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4628 if (Scalar->isVoidTy() || VF == 1)
4630 return VectorType::get(Scalar, VF);
4633 char LoopVectorize::ID = 0;
4634 static const char lv_name[] = "Loop Vectorization";
4635 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4636 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4637 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4638 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4639 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4642 Pass *createLoopVectorizePass() {
4643 return new LoopVectorize();
4647 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4648 // Check for a store.
4649 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4650 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4652 // Check for a load.
4653 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4654 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;