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 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/StringExtras.h"
59 #include "llvm/Analysis/AliasAnalysis.h"
60 #include "llvm/Analysis/BasicAliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/DemandedBits.h"
66 #include "llvm/Analysis/GlobalsModRef.h"
67 #include "llvm/Analysis/LoopAccessAnalysis.h"
68 #include "llvm/Analysis/LoopInfo.h"
69 #include "llvm/Analysis/LoopIterator.h"
70 #include "llvm/Analysis/LoopPass.h"
71 #include "llvm/Analysis/ScalarEvolution.h"
72 #include "llvm/Analysis/ScalarEvolutionExpander.h"
73 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
74 #include "llvm/Analysis/TargetTransformInfo.h"
75 #include "llvm/Analysis/ValueTracking.h"
76 #include "llvm/IR/Constants.h"
77 #include "llvm/IR/DataLayout.h"
78 #include "llvm/IR/DebugInfo.h"
79 #include "llvm/IR/DerivedTypes.h"
80 #include "llvm/IR/DiagnosticInfo.h"
81 #include "llvm/IR/Dominators.h"
82 #include "llvm/IR/Function.h"
83 #include "llvm/IR/IRBuilder.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/LLVMContext.h"
87 #include "llvm/IR/Module.h"
88 #include "llvm/IR/PatternMatch.h"
89 #include "llvm/IR/Type.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/IR/Verifier.h"
93 #include "llvm/Pass.h"
94 #include "llvm/Support/BranchProbability.h"
95 #include "llvm/Support/CommandLine.h"
96 #include "llvm/Support/Debug.h"
97 #include "llvm/Support/raw_ostream.h"
98 #include "llvm/Transforms/Scalar.h"
99 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
100 #include "llvm/Transforms/Utils/Local.h"
101 #include "llvm/Analysis/VectorUtils.h"
102 #include "llvm/Transforms/Utils/LoopUtils.h"
104 #include <functional>
108 using namespace llvm;
109 using namespace llvm::PatternMatch;
111 #define LV_NAME "loop-vectorize"
112 #define DEBUG_TYPE LV_NAME
114 STATISTIC(LoopsVectorized, "Number of loops vectorized");
115 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
129 /// This enables versioning on the strides of symbolically striding memory
130 /// accesses in code like the following.
131 /// for (i = 0; i < N; ++i)
132 /// A[i * Stride1] += B[i * Stride2] ...
134 /// Will be roughly translated to
135 /// if (Stride1 == 1 && Stride2 == 1) {
136 /// for (i = 0; i < N; i+=4)
140 static cl::opt<bool> EnableMemAccessVersioning(
141 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
142 cl::desc("Enable symblic stride memory access versioning"));
144 static cl::opt<bool> EnableInterleavedMemAccesses(
145 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
146 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
148 /// Maximum factor for an interleaved memory access.
149 static cl::opt<unsigned> MaxInterleaveGroupFactor(
150 "max-interleave-group-factor", cl::Hidden,
151 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
154 /// We don't interleave loops with a known constant trip count below this
156 static const unsigned TinyTripCountInterleaveThreshold = 128;
158 static cl::opt<unsigned> ForceTargetNumScalarRegs(
159 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of scalar registers."));
162 static cl::opt<unsigned> ForceTargetNumVectorRegs(
163 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
164 cl::desc("A flag that overrides the target's number of vector registers."));
166 /// Maximum vectorization interleave count.
167 static const unsigned MaxInterleaveFactor = 16;
169 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
170 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
171 cl::desc("A flag that overrides the target's max interleave factor for "
174 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
175 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
176 cl::desc("A flag that overrides the target's max interleave factor for "
177 "vectorized loops."));
179 static cl::opt<unsigned> ForceTargetInstructionCost(
180 "force-target-instruction-cost", cl::init(0), cl::Hidden,
181 cl::desc("A flag that overrides the target's expected cost for "
182 "an instruction to a single constant value. Mostly "
183 "useful for getting consistent testing."));
185 static cl::opt<unsigned> SmallLoopCost(
186 "small-loop-cost", cl::init(20), cl::Hidden,
188 "The cost of a loop that is considered 'small' by the interleaver."));
190 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
191 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
192 cl::desc("Enable the use of the block frequency analysis to access PGO "
193 "heuristics minimizing code growth in cold regions and being more "
194 "aggressive in hot regions."));
196 // Runtime interleave loops for load/store throughput.
197 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
198 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
200 "Enable runtime interleaving until load/store ports are saturated"));
202 /// The number of stores in a loop that are allowed to need predication.
203 static cl::opt<unsigned> NumberOfStoresToPredicate(
204 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
205 cl::desc("Max number of stores to be predicated behind an if."));
207 static cl::opt<bool> EnableIndVarRegisterHeur(
208 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
209 cl::desc("Count the induction variable only once when interleaving"));
211 static cl::opt<bool> EnableCondStoresVectorization(
212 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
213 cl::desc("Enable if predication of stores during vectorization."));
215 static cl::opt<unsigned> MaxNestedScalarReductionIC(
216 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
217 cl::desc("The maximum interleave count to use when interleaving a scalar "
218 "reduction in a nested loop."));
220 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
221 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
222 cl::desc("The maximum allowed number of runtime memory checks with a "
223 "vectorize(enable) pragma."));
227 // Forward declarations.
228 class LoopVectorizeHints;
229 class LoopVectorizationLegality;
230 class LoopVectorizationCostModel;
231 class LoopVectorizationRequirements;
233 /// \brief This modifies LoopAccessReport to initialize message with
234 /// loop-vectorizer-specific part.
235 class VectorizationReport : public LoopAccessReport {
237 VectorizationReport(Instruction *I = nullptr)
238 : LoopAccessReport("loop not vectorized: ", I) {}
240 /// \brief This allows promotion of the loop-access analysis report into the
241 /// loop-vectorizer report. It modifies the message to add the
242 /// loop-vectorizer-specific part of the message.
243 explicit VectorizationReport(const LoopAccessReport &R)
244 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
248 /// A helper function for converting Scalar types to vector types.
249 /// If the incoming type is void, we return void. If the VF is 1, we return
251 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
252 if (Scalar->isVoidTy() || VF == 1)
254 return VectorType::get(Scalar, VF);
257 /// A helper function that returns GEP instruction and knows to skip
259 static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
261 if (isa<GetElementPtrInst>(Ptr))
262 return cast<GetElementPtrInst>(Ptr);
264 if (isa<BitCastInst>(Ptr) &&
265 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0)))
266 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
270 /// InnerLoopVectorizer vectorizes loops which contain only one basic
271 /// block to a specified vectorization factor (VF).
272 /// This class performs the widening of scalars into vectors, or multiple
273 /// scalars. This class also implements the following features:
274 /// * It inserts an epilogue loop for handling loops that don't have iteration
275 /// counts that are known to be a multiple of the vectorization factor.
276 /// * It handles the code generation for reduction variables.
277 /// * Scalarization (implementation using scalars) of un-vectorizable
279 /// InnerLoopVectorizer does not perform any vectorization-legality
280 /// checks, and relies on the caller to check for the different legality
281 /// aspects. The InnerLoopVectorizer relies on the
282 /// LoopVectorizationLegality class to provide information about the induction
283 /// and reduction variables that were found to a given vectorization factor.
284 class InnerLoopVectorizer {
286 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
287 DominatorTree *DT, const TargetLibraryInfo *TLI,
288 const TargetTransformInfo *TTI, unsigned VecWidth,
289 unsigned UnrollFactor)
290 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
291 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
292 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
293 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr),
294 AddedSafetyChecks(false) {}
296 // Perform the actual loop widening (vectorization).
297 // MinimumBitWidths maps scalar integer values to the smallest bitwidth they
298 // can be validly truncated to. The cost model has assumed this truncation
299 // will happen when vectorizing.
300 void vectorize(LoopVectorizationLegality *L,
301 DenseMap<Instruction*,uint64_t> MinimumBitWidths) {
302 MinBWs = MinimumBitWidths;
304 // Create a new empty loop. Unlink the old loop and connect the new one.
306 // Widen each instruction in the old loop to a new one in the new loop.
307 // Use the Legality module to find the induction and reduction variables.
311 // Return true if any runtime check is added.
312 bool IsSafetyChecksAdded() {
313 return AddedSafetyChecks;
316 virtual ~InnerLoopVectorizer() {}
319 /// A small list of PHINodes.
320 typedef SmallVector<PHINode*, 4> PhiVector;
321 /// When we unroll loops we have multiple vector values for each scalar.
322 /// This data structure holds the unrolled and vectorized values that
323 /// originated from one scalar instruction.
324 typedef SmallVector<Value*, 2> VectorParts;
326 // When we if-convert we need to create edge masks. We have to cache values
327 // so that we don't end up with exponential recursion/IR.
328 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
329 VectorParts> EdgeMaskCache;
331 /// \brief Add checks for strides that were assumed to be 1.
333 /// Returns the last check instruction and the first check instruction in the
334 /// pair as (first, last).
335 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
337 /// Create an empty loop, based on the loop ranges of the old loop.
338 void createEmptyLoop();
339 /// Create a new induction variable inside L.
340 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
341 Value *Step, Instruction *DL);
342 /// Copy and widen the instructions from the old loop.
343 virtual void vectorizeLoop();
345 /// \brief The Loop exit block may have single value PHI nodes where the
346 /// incoming value is 'Undef'. While vectorizing we only handled real values
347 /// that were defined inside the loop. Here we fix the 'undef case'.
351 /// Shrinks vector element sizes based on information in "MinBWs".
352 void truncateToMinimalBitwidths();
354 /// A helper function that computes the predicate of the block BB, assuming
355 /// that the header block of the loop is set to True. It returns the *entry*
356 /// mask for the block BB.
357 VectorParts createBlockInMask(BasicBlock *BB);
358 /// A helper function that computes the predicate of the edge between SRC
360 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
362 /// A helper function to vectorize a single BB within the innermost loop.
363 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
365 /// Vectorize a single PHINode in a block. This method handles the induction
366 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
367 /// arbitrary length vectors.
368 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
369 unsigned UF, unsigned VF, PhiVector *PV);
371 /// Insert the new loop to the loop hierarchy and pass manager
372 /// and update the analysis passes.
373 void updateAnalysis();
375 /// This instruction is un-vectorizable. Implement it as a sequence
376 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
377 /// scalarized instruction behind an if block predicated on the control
378 /// dependence of the instruction.
379 virtual void scalarizeInstruction(Instruction *Instr,
380 bool IfPredicateStore=false);
382 /// Vectorize Load and Store instructions,
383 virtual void vectorizeMemoryInstruction(Instruction *Instr);
385 /// Create a broadcast instruction. This method generates a broadcast
386 /// instruction (shuffle) for loop invariant values and for the induction
387 /// value. If this is the induction variable then we extend it to N, N+1, ...
388 /// this is needed because each iteration in the loop corresponds to a SIMD
390 virtual Value *getBroadcastInstrs(Value *V);
392 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
393 /// to each vector element of Val. The sequence starts at StartIndex.
394 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
396 /// When we go over instructions in the basic block we rely on previous
397 /// values within the current basic block or on loop invariant values.
398 /// When we widen (vectorize) values we place them in the map. If the values
399 /// are not within the map, they have to be loop invariant, so we simply
400 /// broadcast them into a vector.
401 VectorParts &getVectorValue(Value *V);
403 /// Try to vectorize the interleaved access group that \p Instr belongs to.
404 void vectorizeInterleaveGroup(Instruction *Instr);
406 /// Generate a shuffle sequence that will reverse the vector Vec.
407 virtual Value *reverseVector(Value *Vec);
409 /// Returns (and creates if needed) the original loop trip count.
410 Value *getOrCreateTripCount(Loop *NewLoop);
412 /// Returns (and creates if needed) the trip count of the widened loop.
413 Value *getOrCreateVectorTripCount(Loop *NewLoop);
415 /// Emit a bypass check to see if the trip count would overflow, or we
416 /// wouldn't have enough iterations to execute one vector loop.
417 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
418 /// Emit a bypass check to see if the vector trip count is nonzero.
419 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
420 /// Emit bypass checks to check if strides we've assumed to be one really are.
421 void emitStrideChecks(Loop *L, BasicBlock *Bypass);
422 /// Emit bypass checks to check any memory assumptions we may have made.
423 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
425 /// This is a helper class that holds the vectorizer state. It maps scalar
426 /// instructions to vector instructions. When the code is 'unrolled' then
427 /// then a single scalar value is mapped to multiple vector parts. The parts
428 /// are stored in the VectorPart type.
430 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
432 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
434 /// \return True if 'Key' is saved in the Value Map.
435 bool has(Value *Key) const { return MapStorage.count(Key); }
437 /// Initializes a new entry in the map. Sets all of the vector parts to the
438 /// save value in 'Val'.
439 /// \return A reference to a vector with splat values.
440 VectorParts &splat(Value *Key, Value *Val) {
441 VectorParts &Entry = MapStorage[Key];
442 Entry.assign(UF, Val);
446 ///\return A reference to the value that is stored at 'Key'.
447 VectorParts &get(Value *Key) {
448 VectorParts &Entry = MapStorage[Key];
451 assert(Entry.size() == UF);
456 /// The unroll factor. Each entry in the map stores this number of vector
460 /// Map storage. We use std::map and not DenseMap because insertions to a
461 /// dense map invalidates its iterators.
462 std::map<Value *, VectorParts> MapStorage;
465 /// The original loop.
467 /// Scev analysis to use.
475 /// Target Library Info.
476 const TargetLibraryInfo *TLI;
477 /// Target Transform Info.
478 const TargetTransformInfo *TTI;
480 /// The vectorization SIMD factor to use. Each vector will have this many
485 /// The vectorization unroll factor to use. Each scalar is vectorized to this
486 /// many different vector instructions.
489 /// The builder that we use
492 // --- Vectorization state ---
494 /// The vector-loop preheader.
495 BasicBlock *LoopVectorPreHeader;
496 /// The scalar-loop preheader.
497 BasicBlock *LoopScalarPreHeader;
498 /// Middle Block between the vector and the scalar.
499 BasicBlock *LoopMiddleBlock;
500 ///The ExitBlock of the scalar loop.
501 BasicBlock *LoopExitBlock;
502 ///The vector loop body.
503 SmallVector<BasicBlock *, 4> LoopVectorBody;
504 ///The scalar loop body.
505 BasicBlock *LoopScalarBody;
506 /// A list of all bypass blocks. The first block is the entry of the loop.
507 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
509 /// The new Induction variable which was added to the new block.
511 /// The induction variable of the old basic block.
512 PHINode *OldInduction;
513 /// Maps scalars to widened vectors.
515 /// Store instructions that should be predicated, as a pair
516 /// <StoreInst, Predicate>
517 SmallVector<std::pair<StoreInst*,Value*>, 4> PredicatedStores;
518 EdgeMaskCache MaskCache;
519 /// Trip count of the original loop.
521 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
522 Value *VectorTripCount;
524 /// Map of scalar integer values to the smallest bitwidth they can be legally
525 /// represented as. The vector equivalents of these values should be truncated
527 DenseMap<Instruction*,uint64_t> MinBWs;
528 LoopVectorizationLegality *Legal;
530 // Record whether runtime check is added.
531 bool AddedSafetyChecks;
534 class InnerLoopUnroller : public InnerLoopVectorizer {
536 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
537 DominatorTree *DT, const TargetLibraryInfo *TLI,
538 const TargetTransformInfo *TTI, unsigned UnrollFactor)
539 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
542 void scalarizeInstruction(Instruction *Instr,
543 bool IfPredicateStore = false) override;
544 void vectorizeMemoryInstruction(Instruction *Instr) override;
545 Value *getBroadcastInstrs(Value *V) override;
546 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
547 Value *reverseVector(Value *Vec) override;
550 /// \brief Look for a meaningful debug location on the instruction or it's
552 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
557 if (I->getDebugLoc() != Empty)
560 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
561 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
562 if (OpInst->getDebugLoc() != Empty)
569 /// \brief Set the debug location in the builder using the debug location in the
571 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
572 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
573 B.SetCurrentDebugLocation(Inst->getDebugLoc());
575 B.SetCurrentDebugLocation(DebugLoc());
579 /// \return string containing a file name and a line # for the given loop.
580 static std::string getDebugLocString(const Loop *L) {
583 raw_string_ostream OS(Result);
584 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
585 LoopDbgLoc.print(OS);
587 // Just print the module name.
588 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
595 /// \brief Propagate known metadata from one instruction to another.
596 static void propagateMetadata(Instruction *To, const Instruction *From) {
597 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
598 From->getAllMetadataOtherThanDebugLoc(Metadata);
600 for (auto M : Metadata) {
601 unsigned Kind = M.first;
603 // These are safe to transfer (this is safe for TBAA, even when we
604 // if-convert, because should that metadata have had a control dependency
605 // on the condition, and thus actually aliased with some other
606 // non-speculated memory access when the condition was false, this would be
607 // caught by the runtime overlap checks).
608 if (Kind != LLVMContext::MD_tbaa &&
609 Kind != LLVMContext::MD_alias_scope &&
610 Kind != LLVMContext::MD_noalias &&
611 Kind != LLVMContext::MD_fpmath &&
612 Kind != LLVMContext::MD_nontemporal)
615 To->setMetadata(Kind, M.second);
619 /// \brief Propagate known metadata from one instruction to a vector of others.
620 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
622 if (Instruction *I = dyn_cast<Instruction>(V))
623 propagateMetadata(I, From);
626 /// \brief The group of interleaved loads/stores sharing the same stride and
627 /// close to each other.
629 /// Each member in this group has an index starting from 0, and the largest
630 /// index should be less than interleaved factor, which is equal to the absolute
631 /// value of the access's stride.
633 /// E.g. An interleaved load group of factor 4:
634 /// for (unsigned i = 0; i < 1024; i+=4) {
635 /// a = A[i]; // Member of index 0
636 /// b = A[i+1]; // Member of index 1
637 /// d = A[i+3]; // Member of index 3
641 /// An interleaved store group of factor 4:
642 /// for (unsigned i = 0; i < 1024; i+=4) {
644 /// A[i] = a; // Member of index 0
645 /// A[i+1] = b; // Member of index 1
646 /// A[i+2] = c; // Member of index 2
647 /// A[i+3] = d; // Member of index 3
650 /// Note: the interleaved load group could have gaps (missing members), but
651 /// the interleaved store group doesn't allow gaps.
652 class InterleaveGroup {
654 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
655 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
656 assert(Align && "The alignment should be non-zero");
658 Factor = std::abs(Stride);
659 assert(Factor > 1 && "Invalid interleave factor");
661 Reverse = Stride < 0;
665 bool isReverse() const { return Reverse; }
666 unsigned getFactor() const { return Factor; }
667 unsigned getAlignment() const { return Align; }
668 unsigned getNumMembers() const { return Members.size(); }
670 /// \brief Try to insert a new member \p Instr with index \p Index and
671 /// alignment \p NewAlign. The index is related to the leader and it could be
672 /// negative if it is the new leader.
674 /// \returns false if the instruction doesn't belong to the group.
675 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
676 assert(NewAlign && "The new member's alignment should be non-zero");
678 int Key = Index + SmallestKey;
680 // Skip if there is already a member with the same index.
681 if (Members.count(Key))
684 if (Key > LargestKey) {
685 // The largest index is always less than the interleave factor.
686 if (Index >= static_cast<int>(Factor))
690 } else if (Key < SmallestKey) {
691 // The largest index is always less than the interleave factor.
692 if (LargestKey - Key >= static_cast<int>(Factor))
698 // It's always safe to select the minimum alignment.
699 Align = std::min(Align, NewAlign);
700 Members[Key] = Instr;
704 /// \brief Get the member with the given index \p Index
706 /// \returns nullptr if contains no such member.
707 Instruction *getMember(unsigned Index) const {
708 int Key = SmallestKey + Index;
709 if (!Members.count(Key))
712 return Members.find(Key)->second;
715 /// \brief Get the index for the given member. Unlike the key in the member
716 /// map, the index starts from 0.
717 unsigned getIndex(Instruction *Instr) const {
718 for (auto I : Members)
719 if (I.second == Instr)
720 return I.first - SmallestKey;
722 llvm_unreachable("InterleaveGroup contains no such member");
725 Instruction *getInsertPos() const { return InsertPos; }
726 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
729 unsigned Factor; // Interleave Factor.
732 DenseMap<int, Instruction *> Members;
736 // To avoid breaking dependences, vectorized instructions of an interleave
737 // group should be inserted at either the first load or the last store in
740 // E.g. %even = load i32 // Insert Position
741 // %add = add i32 %even // Use of %even
745 // %odd = add i32 // Def of %odd
746 // store i32 %odd // Insert Position
747 Instruction *InsertPos;
750 /// \brief Drive the analysis of interleaved memory accesses in the loop.
752 /// Use this class to analyze interleaved accesses only when we can vectorize
753 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
754 /// on interleaved accesses is unsafe.
756 /// The analysis collects interleave groups and records the relationships
757 /// between the member and the group in a map.
758 class InterleavedAccessInfo {
760 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
761 : SE(SE), TheLoop(L), DT(DT) {}
763 ~InterleavedAccessInfo() {
764 SmallSet<InterleaveGroup *, 4> DelSet;
765 // Avoid releasing a pointer twice.
766 for (auto &I : InterleaveGroupMap)
767 DelSet.insert(I.second);
768 for (auto *Ptr : DelSet)
772 /// \brief Analyze the interleaved accesses and collect them in interleave
773 /// groups. Substitute symbolic strides using \p Strides.
774 void analyzeInterleaving(const ValueToValueMap &Strides);
776 /// \brief Check if \p Instr belongs to any interleave group.
777 bool isInterleaved(Instruction *Instr) const {
778 return InterleaveGroupMap.count(Instr);
781 /// \brief Get the interleave group that \p Instr belongs to.
783 /// \returns nullptr if doesn't have such group.
784 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
785 if (InterleaveGroupMap.count(Instr))
786 return InterleaveGroupMap.find(Instr)->second;
795 /// Holds the relationships between the members and the interleave group.
796 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
798 /// \brief The descriptor for a strided memory access.
799 struct StrideDescriptor {
800 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
802 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
804 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
806 int Stride; // The access's stride. It is negative for a reverse access.
807 const SCEV *Scev; // The scalar expression of this access
808 unsigned Size; // The size of the memory object.
809 unsigned Align; // The alignment of this access.
812 /// \brief Create a new interleave group with the given instruction \p Instr,
813 /// stride \p Stride and alignment \p Align.
815 /// \returns the newly created interleave group.
816 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
818 assert(!InterleaveGroupMap.count(Instr) &&
819 "Already in an interleaved access group");
820 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
821 return InterleaveGroupMap[Instr];
824 /// \brief Release the group and remove all the relationships.
825 void releaseGroup(InterleaveGroup *Group) {
826 for (unsigned i = 0; i < Group->getFactor(); i++)
827 if (Instruction *Member = Group->getMember(i))
828 InterleaveGroupMap.erase(Member);
833 /// \brief Collect all the accesses with a constant stride in program order.
834 void collectConstStridedAccesses(
835 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
836 const ValueToValueMap &Strides);
839 /// Utility class for getting and setting loop vectorizer hints in the form
840 /// of loop metadata.
841 /// This class keeps a number of loop annotations locally (as member variables)
842 /// and can, upon request, write them back as metadata on the loop. It will
843 /// initially scan the loop for existing metadata, and will update the local
844 /// values based on information in the loop.
845 /// We cannot write all values to metadata, as the mere presence of some info,
846 /// for example 'force', means a decision has been made. So, we need to be
847 /// careful NOT to add them if the user hasn't specifically asked so.
848 class LoopVectorizeHints {
855 /// Hint - associates name and validation with the hint value.
858 unsigned Value; // This may have to change for non-numeric values.
861 Hint(const char * Name, unsigned Value, HintKind Kind)
862 : Name(Name), Value(Value), Kind(Kind) { }
864 bool validate(unsigned Val) {
867 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
869 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
877 /// Vectorization width.
879 /// Vectorization interleave factor.
881 /// Vectorization forced
884 /// Return the loop metadata prefix.
885 static StringRef Prefix() { return "llvm.loop."; }
889 FK_Undefined = -1, ///< Not selected.
890 FK_Disabled = 0, ///< Forcing disabled.
891 FK_Enabled = 1, ///< Forcing enabled.
894 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
895 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
897 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
898 Force("vectorize.enable", FK_Undefined, HK_FORCE),
900 // Populate values with existing loop metadata.
901 getHintsFromMetadata();
903 // force-vector-interleave overrides DisableInterleaving.
904 if (VectorizerParams::isInterleaveForced())
905 Interleave.Value = VectorizerParams::VectorizationInterleave;
907 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
908 << "LV: Interleaving disabled by the pass manager\n");
911 /// Mark the loop L as already vectorized by setting the width to 1.
912 void setAlreadyVectorized() {
913 Width.Value = Interleave.Value = 1;
914 Hint Hints[] = {Width, Interleave};
915 writeHintsToMetadata(Hints);
918 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
919 if (getForce() == LoopVectorizeHints::FK_Disabled) {
920 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
921 emitOptimizationRemarkAnalysis(F->getContext(),
922 vectorizeAnalysisPassName(), *F,
923 L->getStartLoc(), emitRemark());
927 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
928 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
929 emitOptimizationRemarkAnalysis(F->getContext(),
930 vectorizeAnalysisPassName(), *F,
931 L->getStartLoc(), emitRemark());
935 if (getWidth() == 1 && getInterleave() == 1) {
936 // FIXME: Add a separate metadata to indicate when the loop has already
937 // been vectorized instead of setting width and count to 1.
938 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
939 // FIXME: Add interleave.disable metadata. This will allow
940 // vectorize.disable to be used without disabling the pass and errors
941 // to differentiate between disabled vectorization and a width of 1.
942 emitOptimizationRemarkAnalysis(
943 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
944 "loop not vectorized: vectorization and interleaving are explicitly "
945 "disabled, or vectorize width and interleave count are both set to "
953 /// Dumps all the hint information.
954 std::string emitRemark() const {
955 VectorizationReport R;
956 if (Force.Value == LoopVectorizeHints::FK_Disabled)
957 R << "vectorization is explicitly disabled";
959 R << "use -Rpass-analysis=loop-vectorize for more info";
960 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
962 if (Width.Value != 0)
963 R << ", Vector Width=" << Width.Value;
964 if (Interleave.Value != 0)
965 R << ", Interleave Count=" << Interleave.Value;
973 unsigned getWidth() const { return Width.Value; }
974 unsigned getInterleave() const { return Interleave.Value; }
975 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
976 const char *vectorizeAnalysisPassName() const {
977 // If hints are provided that don't disable vectorization use the
978 // AlwaysPrint pass name to force the frontend to print the diagnostic.
981 if (getForce() == LoopVectorizeHints::FK_Disabled)
983 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
985 return DiagnosticInfo::AlwaysPrint;
988 bool allowReordering() const {
989 // When enabling loop hints are provided we allow the vectorizer to change
990 // the order of operations that is given by the scalar loop. This is not
991 // enabled by default because can be unsafe or inefficient. For example,
992 // reordering floating-point operations will change the way round-off
993 // error accumulates in the loop.
994 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
998 /// Find hints specified in the loop metadata and update local values.
999 void getHintsFromMetadata() {
1000 MDNode *LoopID = TheLoop->getLoopID();
1004 // First operand should refer to the loop id itself.
1005 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1006 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1008 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1009 const MDString *S = nullptr;
1010 SmallVector<Metadata *, 4> Args;
1012 // The expected hint is either a MDString or a MDNode with the first
1013 // operand a MDString.
1014 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1015 if (!MD || MD->getNumOperands() == 0)
1017 S = dyn_cast<MDString>(MD->getOperand(0));
1018 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1019 Args.push_back(MD->getOperand(i));
1021 S = dyn_cast<MDString>(LoopID->getOperand(i));
1022 assert(Args.size() == 0 && "too many arguments for MDString");
1028 // Check if the hint starts with the loop metadata prefix.
1029 StringRef Name = S->getString();
1030 if (Args.size() == 1)
1031 setHint(Name, Args[0]);
1035 /// Checks string hint with one operand and set value if valid.
1036 void setHint(StringRef Name, Metadata *Arg) {
1037 if (!Name.startswith(Prefix()))
1039 Name = Name.substr(Prefix().size(), StringRef::npos);
1041 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1043 unsigned Val = C->getZExtValue();
1045 Hint *Hints[] = {&Width, &Interleave, &Force};
1046 for (auto H : Hints) {
1047 if (Name == H->Name) {
1048 if (H->validate(Val))
1051 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1057 /// Create a new hint from name / value pair.
1058 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1059 LLVMContext &Context = TheLoop->getHeader()->getContext();
1060 Metadata *MDs[] = {MDString::get(Context, Name),
1061 ConstantAsMetadata::get(
1062 ConstantInt::get(Type::getInt32Ty(Context), V))};
1063 return MDNode::get(Context, MDs);
1066 /// Matches metadata with hint name.
1067 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1068 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1072 for (auto H : HintTypes)
1073 if (Name->getString().endswith(H.Name))
1078 /// Sets current hints into loop metadata, keeping other values intact.
1079 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1080 if (HintTypes.size() == 0)
1083 // Reserve the first element to LoopID (see below).
1084 SmallVector<Metadata *, 4> MDs(1);
1085 // If the loop already has metadata, then ignore the existing operands.
1086 MDNode *LoopID = TheLoop->getLoopID();
1088 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1089 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1090 // If node in update list, ignore old value.
1091 if (!matchesHintMetadataName(Node, HintTypes))
1092 MDs.push_back(Node);
1096 // Now, add the missing hints.
1097 for (auto H : HintTypes)
1098 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1100 // Replace current metadata node with new one.
1101 LLVMContext &Context = TheLoop->getHeader()->getContext();
1102 MDNode *NewLoopID = MDNode::get(Context, MDs);
1103 // Set operand 0 to refer to the loop id itself.
1104 NewLoopID->replaceOperandWith(0, NewLoopID);
1106 TheLoop->setLoopID(NewLoopID);
1109 /// The loop these hints belong to.
1110 const Loop *TheLoop;
1113 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1114 const LoopVectorizeHints &Hints,
1115 const LoopAccessReport &Message) {
1116 const char *Name = Hints.vectorizeAnalysisPassName();
1117 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1120 static void emitMissedWarning(Function *F, Loop *L,
1121 const LoopVectorizeHints &LH) {
1122 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1125 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1126 if (LH.getWidth() != 1)
1127 emitLoopVectorizeWarning(
1128 F->getContext(), *F, L->getStartLoc(),
1129 "failed explicitly specified loop vectorization");
1130 else if (LH.getInterleave() != 1)
1131 emitLoopInterleaveWarning(
1132 F->getContext(), *F, L->getStartLoc(),
1133 "failed explicitly specified loop interleaving");
1137 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1138 /// to what vectorization factor.
1139 /// This class does not look at the profitability of vectorization, only the
1140 /// legality. This class has two main kinds of checks:
1141 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1142 /// will change the order of memory accesses in a way that will change the
1143 /// correctness of the program.
1144 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1145 /// checks for a number of different conditions, such as the availability of a
1146 /// single induction variable, that all types are supported and vectorize-able,
1147 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1148 /// This class is also used by InnerLoopVectorizer for identifying
1149 /// induction variable and the different reduction variables.
1150 class LoopVectorizationLegality {
1152 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1153 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1154 Function *F, const TargetTransformInfo *TTI,
1155 LoopAccessAnalysis *LAA,
1156 LoopVectorizationRequirements *R,
1157 const LoopVectorizeHints *H)
1158 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1159 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1160 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1161 Requirements(R), Hints(H) {}
1163 /// ReductionList contains the reduction descriptors for all
1164 /// of the reductions that were found in the loop.
1165 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1167 /// InductionList saves induction variables and maps them to the
1168 /// induction descriptor.
1169 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1171 /// Returns true if it is legal to vectorize this loop.
1172 /// This does not mean that it is profitable to vectorize this
1173 /// loop, only that it is legal to do so.
1174 bool canVectorize();
1176 /// Returns the Induction variable.
1177 PHINode *getInduction() { return Induction; }
1179 /// Returns the reduction variables found in the loop.
1180 ReductionList *getReductionVars() { return &Reductions; }
1182 /// Returns the induction variables found in the loop.
1183 InductionList *getInductionVars() { return &Inductions; }
1185 /// Returns the widest induction type.
1186 Type *getWidestInductionType() { return WidestIndTy; }
1188 /// Returns True if V is an induction variable in this loop.
1189 bool isInductionVariable(const Value *V);
1191 /// Return true if the block BB needs to be predicated in order for the loop
1192 /// to be vectorized.
1193 bool blockNeedsPredication(BasicBlock *BB);
1195 /// Check if this pointer is consecutive when vectorizing. This happens
1196 /// when the last index of the GEP is the induction variable, or that the
1197 /// pointer itself is an induction variable.
1198 /// This check allows us to vectorize A[idx] into a wide load/store.
1200 /// 0 - Stride is unknown or non-consecutive.
1201 /// 1 - Address is consecutive.
1202 /// -1 - Address is consecutive, and decreasing.
1203 int isConsecutivePtr(Value *Ptr);
1205 /// Returns true if the value V is uniform within the loop.
1206 bool isUniform(Value *V);
1208 /// Returns true if this instruction will remain scalar after vectorization.
1209 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1211 /// Returns the information that we collected about runtime memory check.
1212 const RuntimePointerChecking *getRuntimePointerChecking() const {
1213 return LAI->getRuntimePointerChecking();
1216 const LoopAccessInfo *getLAI() const {
1220 /// \brief Check if \p Instr belongs to any interleaved access group.
1221 bool isAccessInterleaved(Instruction *Instr) {
1222 return InterleaveInfo.isInterleaved(Instr);
1225 /// \brief Get the interleaved access group that \p Instr belongs to.
1226 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1227 return InterleaveInfo.getInterleaveGroup(Instr);
1230 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1232 bool hasStride(Value *V) { return StrideSet.count(V); }
1233 bool mustCheckStrides() { return !StrideSet.empty(); }
1234 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1235 return StrideSet.begin();
1237 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1239 /// Returns true if the target machine supports masked store operation
1240 /// for the given \p DataType and kind of access to \p Ptr.
1241 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1242 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1244 /// Returns true if the target machine supports masked load operation
1245 /// for the given \p DataType and kind of access to \p Ptr.
1246 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1247 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1249 /// Returns true if vector representation of the instruction \p I
1251 bool isMaskRequired(const Instruction* I) {
1252 return (MaskedOp.count(I) != 0);
1254 unsigned getNumStores() const {
1255 return LAI->getNumStores();
1257 unsigned getNumLoads() const {
1258 return LAI->getNumLoads();
1260 unsigned getNumPredStores() const {
1261 return NumPredStores;
1264 /// Check if a single basic block loop is vectorizable.
1265 /// At this point we know that this is a loop with a constant trip count
1266 /// and we only need to check individual instructions.
1267 bool canVectorizeInstrs();
1269 /// When we vectorize loops we may change the order in which
1270 /// we read and write from memory. This method checks if it is
1271 /// legal to vectorize the code, considering only memory constrains.
1272 /// Returns true if the loop is vectorizable
1273 bool canVectorizeMemory();
1275 /// Return true if we can vectorize this loop using the IF-conversion
1277 bool canVectorizeWithIfConvert();
1279 /// Collect the variables that need to stay uniform after vectorization.
1280 void collectLoopUniforms();
1282 /// Return true if all of the instructions in the block can be speculatively
1283 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1284 /// and we know that we can read from them without segfault.
1285 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1287 /// \brief Collect memory access with loop invariant strides.
1289 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1291 void collectStridedAccess(Value *LoadOrStoreInst);
1293 /// Report an analysis message to assist the user in diagnosing loops that are
1294 /// not vectorized. These are handled as LoopAccessReport rather than
1295 /// VectorizationReport because the << operator of VectorizationReport returns
1296 /// LoopAccessReport.
1297 void emitAnalysis(const LoopAccessReport &Message) const {
1298 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1301 unsigned NumPredStores;
1303 /// The loop that we evaluate.
1306 ScalarEvolution *SE;
1307 /// Target Library Info.
1308 TargetLibraryInfo *TLI;
1310 Function *TheFunction;
1311 /// Target Transform Info
1312 const TargetTransformInfo *TTI;
1315 // LoopAccess analysis.
1316 LoopAccessAnalysis *LAA;
1317 // And the loop-accesses info corresponding to this loop. This pointer is
1318 // null until canVectorizeMemory sets it up.
1319 const LoopAccessInfo *LAI;
1321 /// The interleave access information contains groups of interleaved accesses
1322 /// with the same stride and close to each other.
1323 InterleavedAccessInfo InterleaveInfo;
1325 // --- vectorization state --- //
1327 /// Holds the integer induction variable. This is the counter of the
1330 /// Holds the reduction variables.
1331 ReductionList Reductions;
1332 /// Holds all of the induction variables that we found in the loop.
1333 /// Notice that inductions don't need to start at zero and that induction
1334 /// variables can be pointers.
1335 InductionList Inductions;
1336 /// Holds the widest induction type encountered.
1339 /// Allowed outside users. This holds the reduction
1340 /// vars which can be accessed from outside the loop.
1341 SmallPtrSet<Value*, 4> AllowedExit;
1342 /// This set holds the variables which are known to be uniform after
1344 SmallPtrSet<Instruction*, 4> Uniforms;
1346 /// Can we assume the absence of NaNs.
1347 bool HasFunNoNaNAttr;
1349 /// Vectorization requirements that will go through late-evaluation.
1350 LoopVectorizationRequirements *Requirements;
1352 /// Used to emit an analysis of any legality issues.
1353 const LoopVectorizeHints *Hints;
1355 ValueToValueMap Strides;
1356 SmallPtrSet<Value *, 8> StrideSet;
1358 /// While vectorizing these instructions we have to generate a
1359 /// call to the appropriate masked intrinsic
1360 SmallPtrSet<const Instruction*, 8> MaskedOp;
1363 /// LoopVectorizationCostModel - estimates the expected speedups due to
1365 /// In many cases vectorization is not profitable. This can happen because of
1366 /// a number of reasons. In this class we mainly attempt to predict the
1367 /// expected speedup/slowdowns due to the supported instruction set. We use the
1368 /// TargetTransformInfo to query the different backends for the cost of
1369 /// different operations.
1370 class LoopVectorizationCostModel {
1372 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1373 LoopVectorizationLegality *Legal,
1374 const TargetTransformInfo &TTI,
1375 const TargetLibraryInfo *TLI, DemandedBits *DB,
1376 AssumptionCache *AC,
1377 const Function *F, const LoopVectorizeHints *Hints,
1378 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1379 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1380 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1382 /// Information about vectorization costs
1383 struct VectorizationFactor {
1384 unsigned Width; // Vector width with best cost
1385 unsigned Cost; // Cost of the loop with that width
1387 /// \return The most profitable vectorization factor and the cost of that VF.
1388 /// This method checks every power of two up to VF. If UserVF is not ZERO
1389 /// then this vectorization factor will be selected if vectorization is
1391 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1393 /// \return The size (in bits) of the widest type in the code that
1394 /// needs to be vectorized. We ignore values that remain scalar such as
1395 /// 64 bit loop indices.
1396 unsigned getWidestType();
1398 /// \return The desired interleave count.
1399 /// If interleave count has been specified by metadata it will be returned.
1400 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1401 /// are the selected vectorization factor and the cost of the selected VF.
1402 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1405 /// \return The most profitable unroll factor.
1406 /// This method finds the best unroll-factor based on register pressure and
1407 /// other parameters. VF and LoopCost are the selected vectorization factor
1408 /// and the cost of the selected VF.
1409 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1412 /// \brief A struct that represents some properties of the register usage
1414 struct RegisterUsage {
1415 /// Holds the number of loop invariant values that are used in the loop.
1416 unsigned LoopInvariantRegs;
1417 /// Holds the maximum number of concurrent live intervals in the loop.
1418 unsigned MaxLocalUsers;
1419 /// Holds the number of instructions in the loop.
1420 unsigned NumInstructions;
1423 /// \return information about the register usage of the loop.
1424 RegisterUsage calculateRegisterUsage();
1427 /// Returns the expected execution cost. The unit of the cost does
1428 /// not matter because we use the 'cost' units to compare different
1429 /// vector widths. The cost that is returned is *not* normalized by
1430 /// the factor width.
1431 unsigned expectedCost(unsigned VF);
1433 /// Returns the execution time cost of an instruction for a given vector
1434 /// width. Vector width of one means scalar.
1435 unsigned getInstructionCost(Instruction *I, unsigned VF);
1437 /// Returns whether the instruction is a load or store and will be a emitted
1438 /// as a vector operation.
1439 bool isConsecutiveLoadOrStore(Instruction *I);
1441 /// Report an analysis message to assist the user in diagnosing loops that are
1442 /// not vectorized. These are handled as LoopAccessReport rather than
1443 /// VectorizationReport because the << operator of VectorizationReport returns
1444 /// LoopAccessReport.
1445 void emitAnalysis(const LoopAccessReport &Message) const {
1446 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1450 /// Map of scalar integer values to the smallest bitwidth they can be legally
1451 /// represented as. The vector equivalents of these values should be truncated
1453 DenseMap<Instruction*,uint64_t> MinBWs;
1455 /// The loop that we evaluate.
1458 ScalarEvolution *SE;
1459 /// Loop Info analysis.
1461 /// Vectorization legality.
1462 LoopVectorizationLegality *Legal;
1463 /// Vector target information.
1464 const TargetTransformInfo &TTI;
1465 /// Target Library Info.
1466 const TargetLibraryInfo *TLI;
1467 /// Demanded bits analysis
1469 const Function *TheFunction;
1470 // Loop Vectorize Hint.
1471 const LoopVectorizeHints *Hints;
1472 // Values to ignore in the cost model.
1473 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1476 /// \brief This holds vectorization requirements that must be verified late in
1477 /// the process. The requirements are set by legalize and costmodel. Once
1478 /// vectorization has been determined to be possible and profitable the
1479 /// requirements can be verified by looking for metadata or compiler options.
1480 /// For example, some loops require FP commutativity which is only allowed if
1481 /// vectorization is explicitly specified or if the fast-math compiler option
1482 /// has been provided.
1483 /// Late evaluation of these requirements allows helpful diagnostics to be
1484 /// composed that tells the user what need to be done to vectorize the loop. For
1485 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1486 /// evaluation should be used only when diagnostics can generated that can be
1487 /// followed by a non-expert user.
1488 class LoopVectorizationRequirements {
1490 LoopVectorizationRequirements()
1491 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1493 void addUnsafeAlgebraInst(Instruction *I) {
1494 // First unsafe algebra instruction.
1495 if (!UnsafeAlgebraInst)
1496 UnsafeAlgebraInst = I;
1499 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1501 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1502 const char *Name = Hints.vectorizeAnalysisPassName();
1503 bool Failed = false;
1504 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1505 emitOptimizationRemarkAnalysisFPCommute(
1506 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1507 VectorizationReport() << "cannot prove it is safe to reorder "
1508 "floating-point operations");
1512 // Test if runtime memcheck thresholds are exceeded.
1513 bool PragmaThresholdReached =
1514 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1515 bool ThresholdReached =
1516 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1517 if ((ThresholdReached && !Hints.allowReordering()) ||
1518 PragmaThresholdReached) {
1519 emitOptimizationRemarkAnalysisAliasing(
1520 F->getContext(), Name, *F, L->getStartLoc(),
1521 VectorizationReport()
1522 << "cannot prove it is safe to reorder memory operations");
1523 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1531 unsigned NumRuntimePointerChecks;
1532 Instruction *UnsafeAlgebraInst;
1535 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1537 return V.push_back(&L);
1539 for (Loop *InnerL : L)
1540 addInnerLoop(*InnerL, V);
1543 /// The LoopVectorize Pass.
1544 struct LoopVectorize : public FunctionPass {
1545 /// Pass identification, replacement for typeid
1548 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1550 DisableUnrolling(NoUnrolling),
1551 AlwaysVectorize(AlwaysVectorize) {
1552 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1555 ScalarEvolution *SE;
1557 TargetTransformInfo *TTI;
1559 BlockFrequencyInfo *BFI;
1560 TargetLibraryInfo *TLI;
1563 AssumptionCache *AC;
1564 LoopAccessAnalysis *LAA;
1565 bool DisableUnrolling;
1566 bool AlwaysVectorize;
1568 BlockFrequency ColdEntryFreq;
1570 bool runOnFunction(Function &F) override {
1571 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1572 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1573 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1574 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1575 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1576 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1577 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1578 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1579 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1580 LAA = &getAnalysis<LoopAccessAnalysis>();
1581 DB = &getAnalysis<DemandedBits>();
1583 // Compute some weights outside of the loop over the loops. Compute this
1584 // using a BranchProbability to re-use its scaling math.
1585 const BranchProbability ColdProb(1, 5); // 20%
1586 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1589 // 1. the target claims to have no vector registers, and
1590 // 2. interleaving won't help ILP.
1592 // The second condition is necessary because, even if the target has no
1593 // vector registers, loop vectorization may still enable scalar
1595 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1598 // Build up a worklist of inner-loops to vectorize. This is necessary as
1599 // the act of vectorizing or partially unrolling a loop creates new loops
1600 // and can invalidate iterators across the loops.
1601 SmallVector<Loop *, 8> Worklist;
1604 addInnerLoop(*L, Worklist);
1606 LoopsAnalyzed += Worklist.size();
1608 // Now walk the identified inner loops.
1609 bool Changed = false;
1610 while (!Worklist.empty())
1611 Changed |= processLoop(Worklist.pop_back_val());
1613 // Process each loop nest in the function.
1617 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1618 SmallVector<Metadata *, 4> MDs;
1619 // Reserve first location for self reference to the LoopID metadata node.
1620 MDs.push_back(nullptr);
1621 bool IsUnrollMetadata = false;
1622 MDNode *LoopID = L->getLoopID();
1624 // First find existing loop unrolling disable metadata.
1625 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1626 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1628 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1630 S && S->getString().startswith("llvm.loop.unroll.disable");
1632 MDs.push_back(LoopID->getOperand(i));
1636 if (!IsUnrollMetadata) {
1637 // Add runtime unroll disable metadata.
1638 LLVMContext &Context = L->getHeader()->getContext();
1639 SmallVector<Metadata *, 1> DisableOperands;
1640 DisableOperands.push_back(
1641 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1642 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1643 MDs.push_back(DisableNode);
1644 MDNode *NewLoopID = MDNode::get(Context, MDs);
1645 // Set operand 0 to refer to the loop id itself.
1646 NewLoopID->replaceOperandWith(0, NewLoopID);
1647 L->setLoopID(NewLoopID);
1651 bool processLoop(Loop *L) {
1652 assert(L->empty() && "Only process inner loops.");
1655 const std::string DebugLocStr = getDebugLocString(L);
1658 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1659 << L->getHeader()->getParent()->getName() << "\" from "
1660 << DebugLocStr << "\n");
1662 LoopVectorizeHints Hints(L, DisableUnrolling);
1664 DEBUG(dbgs() << "LV: Loop hints:"
1666 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1668 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1670 : "?")) << " width=" << Hints.getWidth()
1671 << " unroll=" << Hints.getInterleave() << "\n");
1673 // Function containing loop
1674 Function *F = L->getHeader()->getParent();
1676 // Looking at the diagnostic output is the only way to determine if a loop
1677 // was vectorized (other than looking at the IR or machine code), so it
1678 // is important to generate an optimization remark for each loop. Most of
1679 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1680 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1681 // less verbose reporting vectorized loops and unvectorized loops that may
1682 // benefit from vectorization, respectively.
1684 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1685 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1689 // Check the loop for a trip count threshold:
1690 // do not vectorize loops with a tiny trip count.
1691 const unsigned TC = SE->getSmallConstantTripCount(L);
1692 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1693 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1694 << "This loop is not worth vectorizing.");
1695 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1696 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1698 DEBUG(dbgs() << "\n");
1699 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1700 << "vectorization is not beneficial "
1701 "and is not explicitly forced");
1706 // Check if it is legal to vectorize the loop.
1707 LoopVectorizationRequirements Requirements;
1708 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1709 &Requirements, &Hints);
1710 if (!LVL.canVectorize()) {
1711 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1712 emitMissedWarning(F, L, Hints);
1716 // Collect values we want to ignore in the cost model. This includes
1717 // type-promoting instructions we identified during reduction detection.
1718 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1719 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1720 for (auto &Reduction : *LVL.getReductionVars()) {
1721 RecurrenceDescriptor &RedDes = Reduction.second;
1722 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1723 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1726 // Use the cost model.
1727 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, DB, AC, F, &Hints,
1730 // Check the function attributes to find out if this function should be
1731 // optimized for size.
1732 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1735 // Compute the weighted frequency of this loop being executed and see if it
1736 // is less than 20% of the function entry baseline frequency. Note that we
1737 // always have a canonical loop here because we think we *can* vectorize.
1738 // FIXME: This is hidden behind a flag due to pervasive problems with
1739 // exactly what block frequency models.
1740 if (LoopVectorizeWithBlockFrequency) {
1741 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1742 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1743 LoopEntryFreq < ColdEntryFreq)
1747 // Check the function attributes to see if implicit floats are allowed.
1748 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1749 // an integer loop and the vector instructions selected are purely integer
1750 // vector instructions?
1751 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1752 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1753 "attribute is used.\n");
1756 VectorizationReport()
1757 << "loop not vectorized due to NoImplicitFloat attribute");
1758 emitMissedWarning(F, L, Hints);
1762 // Select the optimal vectorization factor.
1763 const LoopVectorizationCostModel::VectorizationFactor VF =
1764 CM.selectVectorizationFactor(OptForSize);
1766 // Select the interleave count.
1767 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1769 // Get user interleave count.
1770 unsigned UserIC = Hints.getInterleave();
1772 // Identify the diagnostic messages that should be produced.
1773 std::string VecDiagMsg, IntDiagMsg;
1774 bool VectorizeLoop = true, InterleaveLoop = true;
1776 if (Requirements.doesNotMeet(F, L, Hints)) {
1777 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1779 emitMissedWarning(F, L, Hints);
1783 if (VF.Width == 1) {
1784 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1786 "the cost-model indicates that vectorization is not beneficial";
1787 VectorizeLoop = false;
1790 if (IC == 1 && UserIC <= 1) {
1791 // Tell the user interleaving is not beneficial.
1792 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1794 "the cost-model indicates that interleaving is not beneficial";
1795 InterleaveLoop = false;
1798 " and is explicitly disabled or interleave count is set to 1";
1799 } else if (IC > 1 && UserIC == 1) {
1800 // Tell the user interleaving is beneficial, but it explicitly disabled.
1802 << "LV: Interleaving is beneficial but is explicitly disabled.");
1803 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1804 "but is explicitly disabled or interleave count is set to 1";
1805 InterleaveLoop = false;
1808 // Override IC if user provided an interleave count.
1809 IC = UserIC > 0 ? UserIC : IC;
1811 // Emit diagnostic messages, if any.
1812 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1813 if (!VectorizeLoop && !InterleaveLoop) {
1814 // Do not vectorize or interleaving the loop.
1815 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1816 L->getStartLoc(), VecDiagMsg);
1817 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1818 L->getStartLoc(), IntDiagMsg);
1820 } else if (!VectorizeLoop && InterleaveLoop) {
1821 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1822 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1823 L->getStartLoc(), VecDiagMsg);
1824 } else if (VectorizeLoop && !InterleaveLoop) {
1825 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1826 << DebugLocStr << '\n');
1827 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1828 L->getStartLoc(), IntDiagMsg);
1829 } else if (VectorizeLoop && InterleaveLoop) {
1830 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1831 << DebugLocStr << '\n');
1832 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1835 if (!VectorizeLoop) {
1836 assert(IC > 1 && "interleave count should not be 1 or 0");
1837 // If we decided that it is not legal to vectorize the loop then
1839 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1840 Unroller.vectorize(&LVL, CM.MinBWs);
1842 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1843 Twine("interleaved loop (interleaved count: ") +
1846 // If we decided that it is *legal* to vectorize the loop then do it.
1847 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1848 LB.vectorize(&LVL, CM.MinBWs);
1851 // Add metadata to disable runtime unrolling scalar loop when there's no
1852 // runtime check about strides and memory. Because at this situation,
1853 // scalar loop is rarely used not worthy to be unrolled.
1854 if (!LB.IsSafetyChecksAdded())
1855 AddRuntimeUnrollDisableMetaData(L);
1857 // Report the vectorization decision.
1858 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1859 Twine("vectorized loop (vectorization width: ") +
1860 Twine(VF.Width) + ", interleaved count: " +
1864 // Mark the loop as already vectorized to avoid vectorizing again.
1865 Hints.setAlreadyVectorized();
1867 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1871 void getAnalysisUsage(AnalysisUsage &AU) const override {
1872 AU.addRequired<AssumptionCacheTracker>();
1873 AU.addRequiredID(LoopSimplifyID);
1874 AU.addRequiredID(LCSSAID);
1875 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1876 AU.addRequired<DominatorTreeWrapperPass>();
1877 AU.addRequired<LoopInfoWrapperPass>();
1878 AU.addRequired<ScalarEvolutionWrapperPass>();
1879 AU.addRequired<TargetTransformInfoWrapperPass>();
1880 AU.addRequired<AAResultsWrapperPass>();
1881 AU.addRequired<LoopAccessAnalysis>();
1882 AU.addRequired<DemandedBits>();
1883 AU.addPreserved<LoopInfoWrapperPass>();
1884 AU.addPreserved<DominatorTreeWrapperPass>();
1885 AU.addPreserved<BasicAAWrapperPass>();
1886 AU.addPreserved<AAResultsWrapperPass>();
1887 AU.addPreserved<GlobalsAAWrapperPass>();
1892 } // end anonymous namespace
1894 //===----------------------------------------------------------------------===//
1895 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1896 // LoopVectorizationCostModel.
1897 //===----------------------------------------------------------------------===//
1899 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1900 // We need to place the broadcast of invariant variables outside the loop.
1901 Instruction *Instr = dyn_cast<Instruction>(V);
1903 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1904 Instr->getParent()) != LoopVectorBody.end());
1905 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1907 // Place the code for broadcasting invariant variables in the new preheader.
1908 IRBuilder<>::InsertPointGuard Guard(Builder);
1910 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1912 // Broadcast the scalar into all locations in the vector.
1913 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1918 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1920 assert(Val->getType()->isVectorTy() && "Must be a vector");
1921 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1922 "Elem must be an integer");
1923 assert(Step->getType() == Val->getType()->getScalarType() &&
1924 "Step has wrong type");
1925 // Create the types.
1926 Type *ITy = Val->getType()->getScalarType();
1927 VectorType *Ty = cast<VectorType>(Val->getType());
1928 int VLen = Ty->getNumElements();
1929 SmallVector<Constant*, 8> Indices;
1931 // Create a vector of consecutive numbers from zero to VF.
1932 for (int i = 0; i < VLen; ++i)
1933 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1935 // Add the consecutive indices to the vector value.
1936 Constant *Cv = ConstantVector::get(Indices);
1937 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1938 Step = Builder.CreateVectorSplat(VLen, Step);
1939 assert(Step->getType() == Val->getType() && "Invalid step vec");
1940 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1941 // which can be found from the original scalar operations.
1942 Step = Builder.CreateMul(Cv, Step);
1943 return Builder.CreateAdd(Val, Step, "induction");
1946 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1947 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1948 // Make sure that the pointer does not point to structs.
1949 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1952 // If this value is a pointer induction variable we know it is consecutive.
1953 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1954 if (Phi && Inductions.count(Phi)) {
1955 InductionDescriptor II = Inductions[Phi];
1956 return II.getConsecutiveDirection();
1959 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
1963 unsigned NumOperands = Gep->getNumOperands();
1964 Value *GpPtr = Gep->getPointerOperand();
1965 // If this GEP value is a consecutive pointer induction variable and all of
1966 // the indices are constant then we know it is consecutive. We can
1967 Phi = dyn_cast<PHINode>(GpPtr);
1968 if (Phi && Inductions.count(Phi)) {
1970 // Make sure that the pointer does not point to structs.
1971 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1972 if (GepPtrType->getElementType()->isAggregateType())
1975 // Make sure that all of the index operands are loop invariant.
1976 for (unsigned i = 1; i < NumOperands; ++i)
1977 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1980 InductionDescriptor II = Inductions[Phi];
1981 return II.getConsecutiveDirection();
1984 unsigned InductionOperand = getGEPInductionOperand(Gep);
1986 // Check that all of the gep indices are uniform except for our induction
1988 for (unsigned i = 0; i != NumOperands; ++i)
1989 if (i != InductionOperand &&
1990 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1993 // We can emit wide load/stores only if the last non-zero index is the
1994 // induction variable.
1995 const SCEV *Last = nullptr;
1996 if (!Strides.count(Gep))
1997 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1999 // Because of the multiplication by a stride we can have a s/zext cast.
2000 // We are going to replace this stride by 1 so the cast is safe to ignore.
2002 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
2003 // %0 = trunc i64 %indvars.iv to i32
2004 // %mul = mul i32 %0, %Stride1
2005 // %idxprom = zext i32 %mul to i64 << Safe cast.
2006 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
2008 Last = replaceSymbolicStrideSCEV(SE, Strides,
2009 Gep->getOperand(InductionOperand), Gep);
2010 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
2012 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2016 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2017 const SCEV *Step = AR->getStepRecurrence(*SE);
2019 // The memory is consecutive because the last index is consecutive
2020 // and all other indices are loop invariant.
2023 if (Step->isAllOnesValue())
2030 bool LoopVectorizationLegality::isUniform(Value *V) {
2031 return LAI->isUniform(V);
2034 InnerLoopVectorizer::VectorParts&
2035 InnerLoopVectorizer::getVectorValue(Value *V) {
2036 assert(V != Induction && "The new induction variable should not be used.");
2037 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2039 // If we have a stride that is replaced by one, do it here.
2040 if (Legal->hasStride(V))
2041 V = ConstantInt::get(V->getType(), 1);
2043 // If we have this scalar in the map, return it.
2044 if (WidenMap.has(V))
2045 return WidenMap.get(V);
2047 // If this scalar is unknown, assume that it is a constant or that it is
2048 // loop invariant. Broadcast V and save the value for future uses.
2049 Value *B = getBroadcastInstrs(V);
2051 return WidenMap.splat(V, B);
2054 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2055 assert(Vec->getType()->isVectorTy() && "Invalid type");
2056 SmallVector<Constant*, 8> ShuffleMask;
2057 for (unsigned i = 0; i < VF; ++i)
2058 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2060 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2061 ConstantVector::get(ShuffleMask),
2065 // Get a mask to interleave \p NumVec vectors into a wide vector.
2066 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2067 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2068 // <0, 4, 1, 5, 2, 6, 3, 7>
2069 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2071 SmallVector<Constant *, 16> Mask;
2072 for (unsigned i = 0; i < VF; i++)
2073 for (unsigned j = 0; j < NumVec; j++)
2074 Mask.push_back(Builder.getInt32(j * VF + i));
2076 return ConstantVector::get(Mask);
2079 // Get the strided mask starting from index \p Start.
2080 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2081 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2082 unsigned Stride, unsigned VF) {
2083 SmallVector<Constant *, 16> Mask;
2084 for (unsigned i = 0; i < VF; i++)
2085 Mask.push_back(Builder.getInt32(Start + i * Stride));
2087 return ConstantVector::get(Mask);
2090 // Get a mask of two parts: The first part consists of sequential integers
2091 // starting from 0, The second part consists of UNDEFs.
2092 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2093 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2094 unsigned NumUndef) {
2095 SmallVector<Constant *, 16> Mask;
2096 for (unsigned i = 0; i < NumInt; i++)
2097 Mask.push_back(Builder.getInt32(i));
2099 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2100 for (unsigned i = 0; i < NumUndef; i++)
2101 Mask.push_back(Undef);
2103 return ConstantVector::get(Mask);
2106 // Concatenate two vectors with the same element type. The 2nd vector should
2107 // not have more elements than the 1st vector. If the 2nd vector has less
2108 // elements, extend it with UNDEFs.
2109 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2111 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2112 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2113 assert(VecTy1 && VecTy2 &&
2114 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2115 "Expect two vectors with the same element type");
2117 unsigned NumElts1 = VecTy1->getNumElements();
2118 unsigned NumElts2 = VecTy2->getNumElements();
2119 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2121 if (NumElts1 > NumElts2) {
2122 // Extend with UNDEFs.
2124 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2125 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2128 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2129 return Builder.CreateShuffleVector(V1, V2, Mask);
2132 // Concatenate vectors in the given list. All vectors have the same type.
2133 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2134 ArrayRef<Value *> InputList) {
2135 unsigned NumVec = InputList.size();
2136 assert(NumVec > 1 && "Should be at least two vectors");
2138 SmallVector<Value *, 8> ResList;
2139 ResList.append(InputList.begin(), InputList.end());
2141 SmallVector<Value *, 8> TmpList;
2142 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2143 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2144 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2145 "Only the last vector may have a different type");
2147 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2150 // Push the last vector if the total number of vectors is odd.
2151 if (NumVec % 2 != 0)
2152 TmpList.push_back(ResList[NumVec - 1]);
2155 NumVec = ResList.size();
2156 } while (NumVec > 1);
2161 // Try to vectorize the interleave group that \p Instr belongs to.
2163 // E.g. Translate following interleaved load group (factor = 3):
2164 // for (i = 0; i < N; i+=3) {
2165 // R = Pic[i]; // Member of index 0
2166 // G = Pic[i+1]; // Member of index 1
2167 // B = Pic[i+2]; // Member of index 2
2168 // ... // do something to R, G, B
2171 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2172 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2173 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2174 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2176 // Or translate following interleaved store group (factor = 3):
2177 // for (i = 0; i < N; i+=3) {
2178 // ... do something to R, G, B
2179 // Pic[i] = R; // Member of index 0
2180 // Pic[i+1] = G; // Member of index 1
2181 // Pic[i+2] = B; // Member of index 2
2184 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2185 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2186 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2187 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2188 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2189 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2190 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2191 assert(Group && "Fail to get an interleaved access group.");
2193 // Skip if current instruction is not the insert position.
2194 if (Instr != Group->getInsertPos())
2197 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2198 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2199 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2201 // Prepare for the vector type of the interleaved load/store.
2202 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2203 unsigned InterleaveFactor = Group->getFactor();
2204 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2205 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2207 // Prepare for the new pointers.
2208 setDebugLocFromInst(Builder, Ptr);
2209 VectorParts &PtrParts = getVectorValue(Ptr);
2210 SmallVector<Value *, 2> NewPtrs;
2211 unsigned Index = Group->getIndex(Instr);
2212 for (unsigned Part = 0; Part < UF; Part++) {
2213 // Extract the pointer for current instruction from the pointer vector. A
2214 // reverse access uses the pointer in the last lane.
2215 Value *NewPtr = Builder.CreateExtractElement(
2217 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2219 // Notice current instruction could be any index. Need to adjust the address
2220 // to the member of index 0.
2222 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2223 // b = A[i]; // Member of index 0
2224 // Current pointer is pointed to A[i+1], adjust it to A[i].
2226 // E.g. A[i+1] = a; // Member of index 1
2227 // A[i] = b; // Member of index 0
2228 // A[i+2] = c; // Member of index 2 (Current instruction)
2229 // Current pointer is pointed to A[i+2], adjust it to A[i].
2230 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2232 // Cast to the vector pointer type.
2233 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2236 setDebugLocFromInst(Builder, Instr);
2237 Value *UndefVec = UndefValue::get(VecTy);
2239 // Vectorize the interleaved load group.
2241 for (unsigned Part = 0; Part < UF; Part++) {
2242 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2243 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2245 for (unsigned i = 0; i < InterleaveFactor; i++) {
2246 Instruction *Member = Group->getMember(i);
2248 // Skip the gaps in the group.
2252 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2253 Value *StridedVec = Builder.CreateShuffleVector(
2254 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2256 // If this member has different type, cast the result type.
2257 if (Member->getType() != ScalarTy) {
2258 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2259 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2262 VectorParts &Entry = WidenMap.get(Member);
2264 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2267 propagateMetadata(NewLoadInstr, Instr);
2272 // The sub vector type for current instruction.
2273 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2275 // Vectorize the interleaved store group.
2276 for (unsigned Part = 0; Part < UF; Part++) {
2277 // Collect the stored vector from each member.
2278 SmallVector<Value *, 4> StoredVecs;
2279 for (unsigned i = 0; i < InterleaveFactor; i++) {
2280 // Interleaved store group doesn't allow a gap, so each index has a member
2281 Instruction *Member = Group->getMember(i);
2282 assert(Member && "Fail to get a member from an interleaved store group");
2285 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2286 if (Group->isReverse())
2287 StoredVec = reverseVector(StoredVec);
2289 // If this member has different type, cast it to an unified type.
2290 if (StoredVec->getType() != SubVT)
2291 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2293 StoredVecs.push_back(StoredVec);
2296 // Concatenate all vectors into a wide vector.
2297 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2299 // Interleave the elements in the wide vector.
2300 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2301 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2304 Instruction *NewStoreInstr =
2305 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2306 propagateMetadata(NewStoreInstr, Instr);
2310 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2311 // Attempt to issue a wide load.
2312 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2313 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2315 assert((LI || SI) && "Invalid Load/Store instruction");
2317 // Try to vectorize the interleave group if this access is interleaved.
2318 if (Legal->isAccessInterleaved(Instr))
2319 return vectorizeInterleaveGroup(Instr);
2321 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2322 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2323 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2324 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2325 // An alignment of 0 means target abi alignment. We need to use the scalar's
2326 // target abi alignment in such a case.
2327 const DataLayout &DL = Instr->getModule()->getDataLayout();
2329 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2330 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2331 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2332 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2334 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2335 !Legal->isMaskRequired(SI))
2336 return scalarizeInstruction(Instr, true);
2338 if (ScalarAllocatedSize != VectorElementSize)
2339 return scalarizeInstruction(Instr);
2341 // If the pointer is loop invariant or if it is non-consecutive,
2342 // scalarize the load.
2343 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2344 bool Reverse = ConsecutiveStride < 0;
2345 bool UniformLoad = LI && Legal->isUniform(Ptr);
2346 if (!ConsecutiveStride || UniformLoad)
2347 return scalarizeInstruction(Instr);
2349 Constant *Zero = Builder.getInt32(0);
2350 VectorParts &Entry = WidenMap.get(Instr);
2352 // Handle consecutive loads/stores.
2353 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2354 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2355 setDebugLocFromInst(Builder, Gep);
2356 Value *PtrOperand = Gep->getPointerOperand();
2357 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2358 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2360 // Create the new GEP with the new induction variable.
2361 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2362 Gep2->setOperand(0, FirstBasePtr);
2363 Gep2->setName("gep.indvar.base");
2364 Ptr = Builder.Insert(Gep2);
2366 setDebugLocFromInst(Builder, Gep);
2367 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2368 OrigLoop) && "Base ptr must be invariant");
2370 // The last index does not have to be the induction. It can be
2371 // consecutive and be a function of the index. For example A[I+1];
2372 unsigned NumOperands = Gep->getNumOperands();
2373 unsigned InductionOperand = getGEPInductionOperand(Gep);
2374 // Create the new GEP with the new induction variable.
2375 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2377 for (unsigned i = 0; i < NumOperands; ++i) {
2378 Value *GepOperand = Gep->getOperand(i);
2379 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2381 // Update last index or loop invariant instruction anchored in loop.
2382 if (i == InductionOperand ||
2383 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2384 assert((i == InductionOperand ||
2385 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2386 "Must be last index or loop invariant");
2388 VectorParts &GEPParts = getVectorValue(GepOperand);
2389 Value *Index = GEPParts[0];
2390 Index = Builder.CreateExtractElement(Index, Zero);
2391 Gep2->setOperand(i, Index);
2392 Gep2->setName("gep.indvar.idx");
2395 Ptr = Builder.Insert(Gep2);
2397 // Use the induction element ptr.
2398 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2399 setDebugLocFromInst(Builder, Ptr);
2400 VectorParts &PtrVal = getVectorValue(Ptr);
2401 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2404 VectorParts Mask = createBlockInMask(Instr->getParent());
2407 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2408 "We do not allow storing to uniform addresses");
2409 setDebugLocFromInst(Builder, SI);
2410 // We don't want to update the value in the map as it might be used in
2411 // another expression. So don't use a reference type for "StoredVal".
2412 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2414 for (unsigned Part = 0; Part < UF; ++Part) {
2415 // Calculate the pointer for the specific unroll-part.
2417 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2420 // If we store to reverse consecutive memory locations, then we need
2421 // to reverse the order of elements in the stored value.
2422 StoredVal[Part] = reverseVector(StoredVal[Part]);
2423 // If the address is consecutive but reversed, then the
2424 // wide store needs to start at the last vector element.
2425 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2426 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2427 Mask[Part] = reverseVector(Mask[Part]);
2430 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2431 DataTy->getPointerTo(AddressSpace));
2434 if (Legal->isMaskRequired(SI))
2435 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2438 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2439 propagateMetadata(NewSI, SI);
2445 assert(LI && "Must have a load instruction");
2446 setDebugLocFromInst(Builder, LI);
2447 for (unsigned Part = 0; Part < UF; ++Part) {
2448 // Calculate the pointer for the specific unroll-part.
2450 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2453 // If the address is consecutive but reversed, then the
2454 // wide load needs to start at the last vector element.
2455 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2456 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2457 Mask[Part] = reverseVector(Mask[Part]);
2461 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2462 DataTy->getPointerTo(AddressSpace));
2463 if (Legal->isMaskRequired(LI))
2464 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2465 UndefValue::get(DataTy),
2466 "wide.masked.load");
2468 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2469 propagateMetadata(NewLI, LI);
2470 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2474 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2475 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2476 // Holds vector parameters or scalars, in case of uniform vals.
2477 SmallVector<VectorParts, 4> Params;
2479 setDebugLocFromInst(Builder, Instr);
2481 // Find all of the vectorized parameters.
2482 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2483 Value *SrcOp = Instr->getOperand(op);
2485 // If we are accessing the old induction variable, use the new one.
2486 if (SrcOp == OldInduction) {
2487 Params.push_back(getVectorValue(SrcOp));
2491 // Try using previously calculated values.
2492 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2494 // If the src is an instruction that appeared earlier in the basic block,
2495 // then it should already be vectorized.
2496 if (SrcInst && OrigLoop->contains(SrcInst)) {
2497 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2498 // The parameter is a vector value from earlier.
2499 Params.push_back(WidenMap.get(SrcInst));
2501 // The parameter is a scalar from outside the loop. Maybe even a constant.
2502 VectorParts Scalars;
2503 Scalars.append(UF, SrcOp);
2504 Params.push_back(Scalars);
2508 assert(Params.size() == Instr->getNumOperands() &&
2509 "Invalid number of operands");
2511 // Does this instruction return a value ?
2512 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2514 Value *UndefVec = IsVoidRetTy ? nullptr :
2515 UndefValue::get(VectorType::get(Instr->getType(), VF));
2516 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2517 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2520 if (IfPredicateStore) {
2521 assert(Instr->getParent()->getSinglePredecessor() &&
2522 "Only support single predecessor blocks");
2523 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2524 Instr->getParent());
2527 // For each vector unroll 'part':
2528 for (unsigned Part = 0; Part < UF; ++Part) {
2529 // For each scalar that we create:
2530 for (unsigned Width = 0; Width < VF; ++Width) {
2533 Value *Cmp = nullptr;
2534 if (IfPredicateStore) {
2535 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2536 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2539 Instruction *Cloned = Instr->clone();
2541 Cloned->setName(Instr->getName() + ".cloned");
2542 // Replace the operands of the cloned instructions with extracted scalars.
2543 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2544 Value *Op = Params[op][Part];
2545 // Param is a vector. Need to extract the right lane.
2546 if (Op->getType()->isVectorTy())
2547 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2548 Cloned->setOperand(op, Op);
2551 // Place the cloned scalar in the new loop.
2552 Builder.Insert(Cloned);
2554 // If the original scalar returns a value we need to place it in a vector
2555 // so that future users will be able to use it.
2557 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2558 Builder.getInt32(Width));
2560 if (IfPredicateStore)
2561 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
2567 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2571 if (Instruction *I = dyn_cast<Instruction>(V))
2572 return I->getParent() == Loc->getParent() ? I : nullptr;
2576 std::pair<Instruction *, Instruction *>
2577 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2578 Instruction *tnullptr = nullptr;
2579 if (!Legal->mustCheckStrides())
2580 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2582 IRBuilder<> ChkBuilder(Loc);
2585 Value *Check = nullptr;
2586 Instruction *FirstInst = nullptr;
2587 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2588 SE = Legal->strides_end();
2590 Value *Ptr = stripIntegerCast(*SI);
2591 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2593 // Store the first instruction we create.
2594 FirstInst = getFirstInst(FirstInst, C, Loc);
2596 Check = ChkBuilder.CreateOr(Check, C);
2601 // We have to do this trickery because the IRBuilder might fold the check to a
2602 // constant expression in which case there is no Instruction anchored in a
2604 LLVMContext &Ctx = Loc->getContext();
2605 Instruction *TheCheck =
2606 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2607 ChkBuilder.Insert(TheCheck, "stride.not.one");
2608 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2610 return std::make_pair(FirstInst, TheCheck);
2613 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L,
2618 BasicBlock *Header = L->getHeader();
2619 BasicBlock *Latch = L->getLoopLatch();
2620 // As we're just creating this loop, it's possible no latch exists
2621 // yet. If so, use the header as this will be a single block loop.
2625 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
2626 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2627 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2629 Builder.SetInsertPoint(Latch->getTerminator());
2631 // Create i+1 and fill the PHINode.
2632 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2633 Induction->addIncoming(Start, L->getLoopPreheader());
2634 Induction->addIncoming(Next, Latch);
2635 // Create the compare.
2636 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2637 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2639 // Now we have two terminators. Remove the old one from the block.
2640 Latch->getTerminator()->eraseFromParent();
2645 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2649 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2650 // Find the loop boundaries.
2651 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(OrigLoop);
2652 assert(BackedgeTakenCount != SE->getCouldNotCompute() && "Invalid loop count");
2654 Type *IdxTy = Legal->getWidestInductionType();
2656 // The exit count might have the type of i64 while the phi is i32. This can
2657 // happen if we have an induction variable that is sign extended before the
2658 // compare. The only way that we get a backedge taken count is that the
2659 // induction variable was signed and as such will not overflow. In such a case
2660 // truncation is legal.
2661 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
2662 IdxTy->getPrimitiveSizeInBits())
2663 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
2664 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
2666 // Get the total trip count from the count by adding 1.
2667 const SCEV *ExitCount = SE->getAddExpr(
2668 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
2670 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2672 // Expand the trip count and place the new instructions in the preheader.
2673 // Notice that the pre-header does not change, only the loop body.
2674 SCEVExpander Exp(*SE, DL, "induction");
2676 // Count holds the overall loop count (N).
2677 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2678 L->getLoopPreheader()->getTerminator());
2680 if (TripCount->getType()->isPointerTy())
2682 CastInst::CreatePointerCast(TripCount, IdxTy,
2683 "exitcount.ptrcnt.to.int",
2684 L->getLoopPreheader()->getTerminator());
2689 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2690 if (VectorTripCount)
2691 return VectorTripCount;
2693 Value *TC = getOrCreateTripCount(L);
2694 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2696 // Now we need to generate the expression for N - (N % VF), which is
2697 // the part that the vectorized body will execute.
2698 // The loop step is equal to the vectorization factor (num of SIMD elements)
2699 // times the unroll factor (num of SIMD instructions).
2700 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2701 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2702 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2704 return VectorTripCount;
2707 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
2708 BasicBlock *Bypass) {
2709 Value *Count = getOrCreateTripCount(L);
2710 BasicBlock *BB = L->getLoopPreheader();
2711 IRBuilder<> Builder(BB->getTerminator());
2713 // Generate code to check that the loop's trip count that we computed by
2714 // adding one to the backedge-taken count will not overflow.
2715 Value *CheckMinIters =
2716 Builder.CreateICmpULT(Count,
2717 ConstantInt::get(Count->getType(), VF * UF),
2720 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2721 "min.iters.checked");
2722 if (L->getParentLoop())
2723 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2724 ReplaceInstWithInst(BB->getTerminator(),
2725 BranchInst::Create(Bypass, NewBB, CheckMinIters));
2726 LoopBypassBlocks.push_back(BB);
2729 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
2730 BasicBlock *Bypass) {
2731 Value *TC = getOrCreateVectorTripCount(L);
2732 BasicBlock *BB = L->getLoopPreheader();
2733 IRBuilder<> Builder(BB->getTerminator());
2735 // Now, compare the new count to zero. If it is zero skip the vector loop and
2736 // jump to the scalar loop.
2737 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
2740 // Generate code to check that the loop's trip count that we computed by
2741 // adding one to the backedge-taken count will not overflow.
2742 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2744 if (L->getParentLoop())
2745 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2746 ReplaceInstWithInst(BB->getTerminator(),
2747 BranchInst::Create(Bypass, NewBB, Cmp));
2748 LoopBypassBlocks.push_back(BB);
2751 void InnerLoopVectorizer::emitStrideChecks(Loop *L,
2752 BasicBlock *Bypass) {
2753 BasicBlock *BB = L->getLoopPreheader();
2755 // Generate the code to check that the strides we assumed to be one are really
2756 // one. We want the new basic block to start at the first instruction in a
2757 // sequence of instructions that form a check.
2758 Instruction *StrideCheck;
2759 Instruction *FirstCheckInst;
2760 std::tie(FirstCheckInst, StrideCheck) = addStrideCheck(BB->getTerminator());
2764 // Create a new block containing the stride check.
2765 BB->setName("vector.stridecheck");
2766 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2767 if (L->getParentLoop())
2768 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2769 ReplaceInstWithInst(BB->getTerminator(),
2770 BranchInst::Create(Bypass, NewBB, StrideCheck));
2771 LoopBypassBlocks.push_back(BB);
2772 AddedSafetyChecks = true;
2775 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L,
2776 BasicBlock *Bypass) {
2777 BasicBlock *BB = L->getLoopPreheader();
2779 // Generate the code that checks in runtime if arrays overlap. We put the
2780 // checks into a separate block to make the more common case of few elements
2782 Instruction *FirstCheckInst;
2783 Instruction *MemRuntimeCheck;
2784 std::tie(FirstCheckInst, MemRuntimeCheck) =
2785 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
2786 if (!MemRuntimeCheck)
2789 // Create a new block containing the memory check.
2790 BB->setName("vector.memcheck");
2791 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2792 if (L->getParentLoop())
2793 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2794 ReplaceInstWithInst(BB->getTerminator(),
2795 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
2796 LoopBypassBlocks.push_back(BB);
2797 AddedSafetyChecks = true;
2801 void InnerLoopVectorizer::createEmptyLoop() {
2803 In this function we generate a new loop. The new loop will contain
2804 the vectorized instructions while the old loop will continue to run the
2807 [ ] <-- loop iteration number check.
2810 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2813 || [ ] <-- vector pre header.
2817 | [ ]_| <-- vector loop.
2820 | -[ ] <--- middle-block.
2823 -|- >[ ] <--- new preheader.
2827 | [ ]_| <-- old scalar loop to handle remainder.
2830 >[ ] <-- exit block.
2834 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2835 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2836 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2837 assert(VectorPH && "Invalid loop structure");
2838 assert(ExitBlock && "Must have an exit block");
2840 // Some loops have a single integer induction variable, while other loops
2841 // don't. One example is c++ iterators that often have multiple pointer
2842 // induction variables. In the code below we also support a case where we
2843 // don't have a single induction variable.
2845 // We try to obtain an induction variable from the original loop as hard
2846 // as possible. However if we don't find one that:
2848 // - counts from zero, stepping by one
2849 // - is the size of the widest induction variable type
2850 // then we create a new one.
2851 OldInduction = Legal->getInduction();
2852 Type *IdxTy = Legal->getWidestInductionType();
2854 // Split the single block loop into the two loop structure described above.
2855 BasicBlock *VecBody =
2856 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2857 BasicBlock *MiddleBlock =
2858 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2859 BasicBlock *ScalarPH =
2860 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2862 // Create and register the new vector loop.
2863 Loop* Lp = new Loop();
2864 Loop *ParentLoop = OrigLoop->getParentLoop();
2866 // Insert the new loop into the loop nest and register the new basic blocks
2867 // before calling any utilities such as SCEV that require valid LoopInfo.
2869 ParentLoop->addChildLoop(Lp);
2870 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2871 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2873 LI->addTopLevelLoop(Lp);
2875 Lp->addBasicBlockToLoop(VecBody, *LI);
2877 // Find the loop boundaries.
2878 Value *Count = getOrCreateTripCount(Lp);
2880 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2882 // We need to test whether the backedge-taken count is uint##_max. Adding one
2883 // to it will cause overflow and an incorrect loop trip count in the vector
2884 // body. In case of overflow we want to directly jump to the scalar remainder
2886 emitMinimumIterationCountCheck(Lp, ScalarPH);
2887 // Now, compare the new count to zero. If it is zero skip the vector loop and
2888 // jump to the scalar loop.
2889 emitVectorLoopEnteredCheck(Lp, ScalarPH);
2890 // Generate the code to check that the strides we assumed to be one are really
2891 // one. We want the new basic block to start at the first instruction in a
2892 // sequence of instructions that form a check.
2893 emitStrideChecks(Lp, ScalarPH);
2894 // Generate the code that checks in runtime if arrays overlap. We put the
2895 // checks into a separate block to make the more common case of few elements
2897 emitMemRuntimeChecks(Lp, ScalarPH);
2899 // Generate the induction variable.
2900 // The loop step is equal to the vectorization factor (num of SIMD elements)
2901 // times the unroll factor (num of SIMD instructions).
2902 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
2903 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2905 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
2906 getDebugLocFromInstOrOperands(OldInduction));
2908 // We are going to resume the execution of the scalar loop.
2909 // Go over all of the induction variables that we found and fix the
2910 // PHIs that are left in the scalar version of the loop.
2911 // The starting values of PHI nodes depend on the counter of the last
2912 // iteration in the vectorized loop.
2913 // If we come from a bypass edge then we need to start from the original
2916 // This variable saves the new starting index for the scalar loop. It is used
2917 // to test if there are any tail iterations left once the vector loop has
2919 LoopVectorizationLegality::InductionList::iterator I, E;
2920 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2921 for (I = List->begin(), E = List->end(); I != E; ++I) {
2922 PHINode *OrigPhi = I->first;
2923 InductionDescriptor II = I->second;
2925 // Create phi nodes to merge from the backedge-taken check block.
2926 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2928 ScalarPH->getTerminator());
2930 if (OrigPhi == OldInduction) {
2931 // We know what the end value is.
2932 EndValue = CountRoundDown;
2934 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
2935 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
2936 II.getStepValue()->getType(),
2938 EndValue = II.transform(B, CRD);
2939 EndValue->setName("ind.end");
2942 // The new PHI merges the original incoming value, in case of a bypass,
2943 // or the value at the end of the vectorized loop.
2944 BCResumeVal->addIncoming(EndValue, MiddleBlock);
2946 // Fix the scalar body counter (PHI node).
2947 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2949 // The old induction's phi node in the scalar body needs the truncated
2951 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2952 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2953 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2956 // Add a check in the middle block to see if we have completed
2957 // all of the iterations in the first vector loop.
2958 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2959 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
2960 CountRoundDown, "cmp.n",
2961 MiddleBlock->getTerminator());
2962 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2963 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2965 // Get ready to start creating new instructions into the vectorized body.
2966 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
2969 LoopVectorPreHeader = Lp->getLoopPreheader();
2970 LoopScalarPreHeader = ScalarPH;
2971 LoopMiddleBlock = MiddleBlock;
2972 LoopExitBlock = ExitBlock;
2973 LoopVectorBody.push_back(VecBody);
2974 LoopScalarBody = OldBasicBlock;
2976 LoopVectorizeHints Hints(Lp, true);
2977 Hints.setAlreadyVectorized();
2981 struct CSEDenseMapInfo {
2982 static bool canHandle(Instruction *I) {
2983 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2984 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2986 static inline Instruction *getEmptyKey() {
2987 return DenseMapInfo<Instruction *>::getEmptyKey();
2989 static inline Instruction *getTombstoneKey() {
2990 return DenseMapInfo<Instruction *>::getTombstoneKey();
2992 static unsigned getHashValue(Instruction *I) {
2993 assert(canHandle(I) && "Unknown instruction!");
2994 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2995 I->value_op_end()));
2997 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2998 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2999 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3001 return LHS->isIdenticalTo(RHS);
3006 /// \brief Check whether this block is a predicated block.
3007 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
3008 /// = ...; " blocks. We start with one vectorized basic block. For every
3009 /// conditional block we split this vectorized block. Therefore, every second
3010 /// block will be a predicated one.
3011 static bool isPredicatedBlock(unsigned BlockNum) {
3012 return BlockNum % 2;
3015 ///\brief Perform cse of induction variable instructions.
3016 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
3017 // Perform simple cse.
3018 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3019 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3020 BasicBlock *BB = BBs[i];
3021 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3022 Instruction *In = &*I++;
3024 if (!CSEDenseMapInfo::canHandle(In))
3027 // Check if we can replace this instruction with any of the
3028 // visited instructions.
3029 if (Instruction *V = CSEMap.lookup(In)) {
3030 In->replaceAllUsesWith(V);
3031 In->eraseFromParent();
3034 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3035 // ...;" blocks for predicated stores. Every second block is a predicated
3037 if (isPredicatedBlock(i))
3045 /// \brief Adds a 'fast' flag to floating point operations.
3046 static Value *addFastMathFlag(Value *V) {
3047 if (isa<FPMathOperator>(V)){
3048 FastMathFlags Flags;
3049 Flags.setUnsafeAlgebra();
3050 cast<Instruction>(V)->setFastMathFlags(Flags);
3055 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3056 /// the result needs to be inserted and/or extracted from vectors.
3057 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3058 const TargetTransformInfo &TTI) {
3062 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3065 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3067 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3069 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3075 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3076 // Return the cost of the instruction, including scalarization overhead if it's
3077 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3078 // i.e. either vector version isn't available, or is too expensive.
3079 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3080 const TargetTransformInfo &TTI,
3081 const TargetLibraryInfo *TLI,
3082 bool &NeedToScalarize) {
3083 Function *F = CI->getCalledFunction();
3084 StringRef FnName = CI->getCalledFunction()->getName();
3085 Type *ScalarRetTy = CI->getType();
3086 SmallVector<Type *, 4> Tys, ScalarTys;
3087 for (auto &ArgOp : CI->arg_operands())
3088 ScalarTys.push_back(ArgOp->getType());
3090 // Estimate cost of scalarized vector call. The source operands are assumed
3091 // to be vectors, so we need to extract individual elements from there,
3092 // execute VF scalar calls, and then gather the result into the vector return
3094 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3096 return ScalarCallCost;
3098 // Compute corresponding vector type for return value and arguments.
3099 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3100 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3101 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3103 // Compute costs of unpacking argument values for the scalar calls and
3104 // packing the return values to a vector.
3105 unsigned ScalarizationCost =
3106 getScalarizationOverhead(RetTy, true, false, TTI);
3107 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3108 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3110 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3112 // If we can't emit a vector call for this function, then the currently found
3113 // cost is the cost we need to return.
3114 NeedToScalarize = true;
3115 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3118 // If the corresponding vector cost is cheaper, return its cost.
3119 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3120 if (VectorCallCost < Cost) {
3121 NeedToScalarize = false;
3122 return VectorCallCost;
3127 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3128 // factor VF. Return the cost of the instruction, including scalarization
3129 // overhead if it's needed.
3130 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3131 const TargetTransformInfo &TTI,
3132 const TargetLibraryInfo *TLI) {
3133 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3134 assert(ID && "Expected intrinsic call!");
3136 Type *RetTy = ToVectorTy(CI->getType(), VF);
3137 SmallVector<Type *, 4> Tys;
3138 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3139 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3141 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3144 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3145 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3146 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3147 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3149 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3150 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3151 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3152 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3155 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3156 // For every instruction `I` in MinBWs, truncate the operands, create a
3157 // truncated version of `I` and reextend its result. InstCombine runs
3158 // later and will remove any ext/trunc pairs.
3160 for (auto &KV : MinBWs) {
3161 VectorParts &Parts = WidenMap.get(KV.first);
3162 for (Value *&I : Parts) {
3165 Type *OriginalTy = I->getType();
3166 Type *ScalarTruncatedTy = IntegerType::get(OriginalTy->getContext(),
3168 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3169 OriginalTy->getVectorNumElements());
3170 if (TruncatedTy == OriginalTy)
3173 IRBuilder<> B(cast<Instruction>(I));
3174 auto ShrinkOperand = [&](Value *V) -> Value* {
3175 if (auto *ZI = dyn_cast<ZExtInst>(V))
3176 if (ZI->getSrcTy() == TruncatedTy)
3177 return ZI->getOperand(0);
3178 return B.CreateZExtOrTrunc(V, TruncatedTy);
3181 // The actual instruction modification depends on the instruction type,
3183 Value *NewI = nullptr;
3184 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3185 NewI = B.CreateBinOp(BO->getOpcode(),
3186 ShrinkOperand(BO->getOperand(0)),
3187 ShrinkOperand(BO->getOperand(1)));
3188 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3189 } else if (ICmpInst *CI = dyn_cast<ICmpInst>(I)) {
3190 NewI = B.CreateICmp(CI->getPredicate(),
3191 ShrinkOperand(CI->getOperand(0)),
3192 ShrinkOperand(CI->getOperand(1)));
3193 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
3194 NewI = B.CreateSelect(SI->getCondition(),
3195 ShrinkOperand(SI->getTrueValue()),
3196 ShrinkOperand(SI->getFalseValue()));
3197 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3198 switch (CI->getOpcode()) {
3199 default: llvm_unreachable("Unhandled cast!");
3200 case Instruction::Trunc:
3201 NewI = ShrinkOperand(CI->getOperand(0));
3203 case Instruction::SExt:
3204 NewI = B.CreateSExtOrTrunc(CI->getOperand(0),
3205 smallestIntegerVectorType(OriginalTy,
3208 case Instruction::ZExt:
3209 NewI = B.CreateZExtOrTrunc(CI->getOperand(0),
3210 smallestIntegerVectorType(OriginalTy,
3214 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
3215 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3217 B.CreateZExtOrTrunc(SI->getOperand(0),
3218 VectorType::get(ScalarTruncatedTy, Elements0));
3219 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3221 B.CreateZExtOrTrunc(SI->getOperand(1),
3222 VectorType::get(ScalarTruncatedTy, Elements1));
3224 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3225 } else if (isa<LoadInst>(I)) {
3226 // Don't do anything with the operands, just extend the result.
3229 llvm_unreachable("Unhandled instruction type!");
3232 // Lastly, extend the result.
3233 NewI->takeName(cast<Instruction>(I));
3234 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3235 I->replaceAllUsesWith(Res);
3236 cast<Instruction>(I)->eraseFromParent();
3241 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3242 for (auto &KV : MinBWs) {
3243 VectorParts &Parts = WidenMap.get(KV.first);
3244 for (Value *&I : Parts) {
3245 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3246 if (Inst && Inst->use_empty()) {
3247 Value *NewI = Inst->getOperand(0);
3248 Inst->eraseFromParent();
3255 void InnerLoopVectorizer::vectorizeLoop() {
3256 //===------------------------------------------------===//
3258 // Notice: any optimization or new instruction that go
3259 // into the code below should be also be implemented in
3262 //===------------------------------------------------===//
3263 Constant *Zero = Builder.getInt32(0);
3265 // In order to support reduction variables we need to be able to vectorize
3266 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3267 // stages. First, we create a new vector PHI node with no incoming edges.
3268 // We use this value when we vectorize all of the instructions that use the
3269 // PHI. Next, after all of the instructions in the block are complete we
3270 // add the new incoming edges to the PHI. At this point all of the
3271 // instructions in the basic block are vectorized, so we can use them to
3272 // construct the PHI.
3273 PhiVector RdxPHIsToFix;
3275 // Scan the loop in a topological order to ensure that defs are vectorized
3277 LoopBlocksDFS DFS(OrigLoop);
3280 // Vectorize all of the blocks in the original loop.
3281 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3282 be = DFS.endRPO(); bb != be; ++bb)
3283 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3285 // Insert truncates and extends for any truncated instructions as hints to
3288 truncateToMinimalBitwidths();
3290 // At this point every instruction in the original loop is widened to
3291 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3292 // that we vectorized. The PHI nodes are currently empty because we did
3293 // not want to introduce cycles. Notice that the remaining PHI nodes
3294 // that we need to fix are reduction variables.
3296 // Create the 'reduced' values for each of the induction vars.
3297 // The reduced values are the vector values that we scalarize and combine
3298 // after the loop is finished.
3299 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3301 PHINode *RdxPhi = *it;
3302 assert(RdxPhi && "Unable to recover vectorized PHI");
3304 // Find the reduction variable descriptor.
3305 assert(Legal->getReductionVars()->count(RdxPhi) &&
3306 "Unable to find the reduction variable");
3307 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3309 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3310 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3311 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3312 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3313 RdxDesc.getMinMaxRecurrenceKind();
3314 setDebugLocFromInst(Builder, ReductionStartValue);
3316 // We need to generate a reduction vector from the incoming scalar.
3317 // To do so, we need to generate the 'identity' vector and override
3318 // one of the elements with the incoming scalar reduction. We need
3319 // to do it in the vector-loop preheader.
3320 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3322 // This is the vector-clone of the value that leaves the loop.
3323 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3324 Type *VecTy = VectorExit[0]->getType();
3326 // Find the reduction identity variable. Zero for addition, or, xor,
3327 // one for multiplication, -1 for And.
3330 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3331 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3332 // MinMax reduction have the start value as their identify.
3334 VectorStart = Identity = ReductionStartValue;
3336 VectorStart = Identity =
3337 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3340 // Handle other reduction kinds:
3341 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3342 RK, VecTy->getScalarType());
3345 // This vector is the Identity vector where the first element is the
3346 // incoming scalar reduction.
3347 VectorStart = ReductionStartValue;
3349 Identity = ConstantVector::getSplat(VF, Iden);
3351 // This vector is the Identity vector where the first element is the
3352 // incoming scalar reduction.
3354 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3358 // Fix the vector-loop phi.
3360 // Reductions do not have to start at zero. They can start with
3361 // any loop invariant values.
3362 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3363 BasicBlock *Latch = OrigLoop->getLoopLatch();
3364 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3365 VectorParts &Val = getVectorValue(LoopVal);
3366 for (unsigned part = 0; part < UF; ++part) {
3367 // Make sure to add the reduction stat value only to the
3368 // first unroll part.
3369 Value *StartVal = (part == 0) ? VectorStart : Identity;
3370 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3371 LoopVectorPreHeader);
3372 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3373 LoopVectorBody.back());
3376 // Before each round, move the insertion point right between
3377 // the PHIs and the values we are going to write.
3378 // This allows us to write both PHINodes and the extractelement
3380 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3382 VectorParts RdxParts = getVectorValue(LoopExitInst);
3383 setDebugLocFromInst(Builder, LoopExitInst);
3385 // If the vector reduction can be performed in a smaller type, we truncate
3386 // then extend the loop exit value to enable InstCombine to evaluate the
3387 // entire expression in the smaller type.
3388 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3389 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3390 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3391 for (unsigned part = 0; part < UF; ++part) {
3392 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3393 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3394 : Builder.CreateZExt(Trunc, VecTy);
3395 for (Value::user_iterator UI = RdxParts[part]->user_begin();
3396 UI != RdxParts[part]->user_end();)
3398 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
3399 RdxParts[part] = Extnd;
3404 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3405 for (unsigned part = 0; part < UF; ++part)
3406 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3409 // Reduce all of the unrolled parts into a single vector.
3410 Value *ReducedPartRdx = RdxParts[0];
3411 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3412 setDebugLocFromInst(Builder, ReducedPartRdx);
3413 for (unsigned part = 1; part < UF; ++part) {
3414 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3415 // Floating point operations had to be 'fast' to enable the reduction.
3416 ReducedPartRdx = addFastMathFlag(
3417 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3418 ReducedPartRdx, "bin.rdx"));
3420 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3421 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3425 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3426 // and vector ops, reducing the set of values being computed by half each
3428 assert(isPowerOf2_32(VF) &&
3429 "Reduction emission only supported for pow2 vectors!");
3430 Value *TmpVec = ReducedPartRdx;
3431 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3432 for (unsigned i = VF; i != 1; i >>= 1) {
3433 // Move the upper half of the vector to the lower half.
3434 for (unsigned j = 0; j != i/2; ++j)
3435 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3437 // Fill the rest of the mask with undef.
3438 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3439 UndefValue::get(Builder.getInt32Ty()));
3442 Builder.CreateShuffleVector(TmpVec,
3443 UndefValue::get(TmpVec->getType()),
3444 ConstantVector::get(ShuffleMask),
3447 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3448 // Floating point operations had to be 'fast' to enable the reduction.
3449 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3450 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3452 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3456 // The result is in the first element of the vector.
3457 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3458 Builder.getInt32(0));
3460 // If the reduction can be performed in a smaller type, we need to extend
3461 // the reduction to the wider type before we branch to the original loop.
3462 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3465 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3466 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3469 // Create a phi node that merges control-flow from the backedge-taken check
3470 // block and the middle block.
3471 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3472 LoopScalarPreHeader->getTerminator());
3473 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
3474 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
3475 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3477 // Now, we need to fix the users of the reduction variable
3478 // inside and outside of the scalar remainder loop.
3479 // We know that the loop is in LCSSA form. We need to update the
3480 // PHI nodes in the exit blocks.
3481 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3482 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3483 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3484 if (!LCSSAPhi) break;
3486 // All PHINodes need to have a single entry edge, or two if
3487 // we already fixed them.
3488 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3490 // We found our reduction value exit-PHI. Update it with the
3491 // incoming bypass edge.
3492 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3493 // Add an edge coming from the bypass.
3494 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3497 }// end of the LCSSA phi scan.
3499 // Fix the scalar loop reduction variable with the incoming reduction sum
3500 // from the vector body and from the backedge value.
3501 int IncomingEdgeBlockIdx =
3502 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3503 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3504 // Pick the other block.
3505 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3506 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3507 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3508 }// end of for each redux variable.
3512 // Make sure DomTree is updated.
3515 // Predicate any stores.
3516 for (auto KV : PredicatedStores) {
3517 BasicBlock::iterator I(KV.first);
3518 auto *BB = SplitBlock(I->getParent(), &*std::next(I), DT, LI);
3519 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
3520 /*BranchWeights=*/nullptr, DT);
3522 I->getParent()->setName("pred.store.if");
3523 BB->setName("pred.store.continue");
3525 DEBUG(DT->verifyDomTree());
3526 // Remove redundant induction instructions.
3527 cse(LoopVectorBody);
3530 void InnerLoopVectorizer::fixLCSSAPHIs() {
3531 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3532 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3533 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3534 if (!LCSSAPhi) break;
3535 if (LCSSAPhi->getNumIncomingValues() == 1)
3536 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3541 InnerLoopVectorizer::VectorParts
3542 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3543 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3546 // Look for cached value.
3547 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3548 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3549 if (ECEntryIt != MaskCache.end())
3550 return ECEntryIt->second;
3552 VectorParts SrcMask = createBlockInMask(Src);
3554 // The terminator has to be a branch inst!
3555 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3556 assert(BI && "Unexpected terminator found");
3558 if (BI->isConditional()) {
3559 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3561 if (BI->getSuccessor(0) != Dst)
3562 for (unsigned part = 0; part < UF; ++part)
3563 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3565 for (unsigned part = 0; part < UF; ++part)
3566 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3568 MaskCache[Edge] = EdgeMask;
3572 MaskCache[Edge] = SrcMask;
3576 InnerLoopVectorizer::VectorParts
3577 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3578 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3580 // Loop incoming mask is all-one.
3581 if (OrigLoop->getHeader() == BB) {
3582 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3583 return getVectorValue(C);
3586 // This is the block mask. We OR all incoming edges, and with zero.
3587 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3588 VectorParts BlockMask = getVectorValue(Zero);
3591 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3592 VectorParts EM = createEdgeMask(*it, BB);
3593 for (unsigned part = 0; part < UF; ++part)
3594 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3600 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3601 InnerLoopVectorizer::VectorParts &Entry,
3602 unsigned UF, unsigned VF, PhiVector *PV) {
3603 PHINode* P = cast<PHINode>(PN);
3604 // Handle reduction variables:
3605 if (Legal->getReductionVars()->count(P)) {
3606 for (unsigned part = 0; part < UF; ++part) {
3607 // This is phase one of vectorizing PHIs.
3608 Type *VecTy = (VF == 1) ? PN->getType() :
3609 VectorType::get(PN->getType(), VF);
3610 Entry[part] = PHINode::Create(
3611 VecTy, 2, "vec.phi", &*LoopVectorBody.back()->getFirstInsertionPt());
3617 setDebugLocFromInst(Builder, P);
3618 // Check for PHI nodes that are lowered to vector selects.
3619 if (P->getParent() != OrigLoop->getHeader()) {
3620 // We know that all PHIs in non-header blocks are converted into
3621 // selects, so we don't have to worry about the insertion order and we
3622 // can just use the builder.
3623 // At this point we generate the predication tree. There may be
3624 // duplications since this is a simple recursive scan, but future
3625 // optimizations will clean it up.
3627 unsigned NumIncoming = P->getNumIncomingValues();
3629 // Generate a sequence of selects of the form:
3630 // SELECT(Mask3, In3,
3631 // SELECT(Mask2, In2,
3633 for (unsigned In = 0; In < NumIncoming; In++) {
3634 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3636 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3638 for (unsigned part = 0; part < UF; ++part) {
3639 // We might have single edge PHIs (blocks) - use an identity
3640 // 'select' for the first PHI operand.
3642 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3645 // Select between the current value and the previous incoming edge
3646 // based on the incoming mask.
3647 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3648 Entry[part], "predphi");
3654 // This PHINode must be an induction variable.
3655 // Make sure that we know about it.
3656 assert(Legal->getInductionVars()->count(P) &&
3657 "Not an induction variable");
3659 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3661 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3662 // which can be found from the original scalar operations.
3663 switch (II.getKind()) {
3664 case InductionDescriptor::IK_NoInduction:
3665 llvm_unreachable("Unknown induction");
3666 case InductionDescriptor::IK_IntInduction: {
3667 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3668 // Handle other induction variables that are now based on the
3670 Value *V = Induction;
3671 if (P != OldInduction) {
3672 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3673 V = II.transform(Builder, V);
3674 V->setName("offset.idx");
3676 Value *Broadcasted = getBroadcastInstrs(V);
3677 // After broadcasting the induction variable we need to make the vector
3678 // consecutive by adding 0, 1, 2, etc.
3679 for (unsigned part = 0; part < UF; ++part)
3680 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3683 case InductionDescriptor::IK_PtrInduction:
3684 // Handle the pointer induction variable case.
3685 assert(P->getType()->isPointerTy() && "Unexpected type.");
3686 // This is the normalized GEP that starts counting at zero.
3687 Value *PtrInd = Induction;
3688 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3689 // This is the vector of results. Notice that we don't generate
3690 // vector geps because scalar geps result in better code.
3691 for (unsigned part = 0; part < UF; ++part) {
3693 int EltIndex = part;
3694 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3695 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3696 Value *SclrGep = II.transform(Builder, GlobalIdx);
3697 SclrGep->setName("next.gep");
3698 Entry[part] = SclrGep;
3702 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3703 for (unsigned int i = 0; i < VF; ++i) {
3704 int EltIndex = i + part * VF;
3705 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3706 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3707 Value *SclrGep = II.transform(Builder, GlobalIdx);
3708 SclrGep->setName("next.gep");
3709 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3710 Builder.getInt32(i),
3713 Entry[part] = VecVal;
3719 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3720 // For each instruction in the old loop.
3721 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3722 VectorParts &Entry = WidenMap.get(&*it);
3724 switch (it->getOpcode()) {
3725 case Instruction::Br:
3726 // Nothing to do for PHIs and BR, since we already took care of the
3727 // loop control flow instructions.
3729 case Instruction::PHI: {
3730 // Vectorize PHINodes.
3731 widenPHIInstruction(&*it, Entry, UF, VF, PV);
3735 case Instruction::Add:
3736 case Instruction::FAdd:
3737 case Instruction::Sub:
3738 case Instruction::FSub:
3739 case Instruction::Mul:
3740 case Instruction::FMul:
3741 case Instruction::UDiv:
3742 case Instruction::SDiv:
3743 case Instruction::FDiv:
3744 case Instruction::URem:
3745 case Instruction::SRem:
3746 case Instruction::FRem:
3747 case Instruction::Shl:
3748 case Instruction::LShr:
3749 case Instruction::AShr:
3750 case Instruction::And:
3751 case Instruction::Or:
3752 case Instruction::Xor: {
3753 // Just widen binops.
3754 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3755 setDebugLocFromInst(Builder, BinOp);
3756 VectorParts &A = getVectorValue(it->getOperand(0));
3757 VectorParts &B = getVectorValue(it->getOperand(1));
3759 // Use this vector value for all users of the original instruction.
3760 for (unsigned Part = 0; Part < UF; ++Part) {
3761 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3763 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3764 VecOp->copyIRFlags(BinOp);
3769 propagateMetadata(Entry, &*it);
3772 case Instruction::Select: {
3774 // If the selector is loop invariant we can create a select
3775 // instruction with a scalar condition. Otherwise, use vector-select.
3776 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3778 setDebugLocFromInst(Builder, &*it);
3780 // The condition can be loop invariant but still defined inside the
3781 // loop. This means that we can't just use the original 'cond' value.
3782 // We have to take the 'vectorized' value and pick the first lane.
3783 // Instcombine will make this a no-op.
3784 VectorParts &Cond = getVectorValue(it->getOperand(0));
3785 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3786 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3788 Value *ScalarCond = (VF == 1) ? Cond[0] :
3789 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3791 for (unsigned Part = 0; Part < UF; ++Part) {
3792 Entry[Part] = Builder.CreateSelect(
3793 InvariantCond ? ScalarCond : Cond[Part],
3798 propagateMetadata(Entry, &*it);
3802 case Instruction::ICmp:
3803 case Instruction::FCmp: {
3804 // Widen compares. Generate vector compares.
3805 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3806 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3807 setDebugLocFromInst(Builder, &*it);
3808 VectorParts &A = getVectorValue(it->getOperand(0));
3809 VectorParts &B = getVectorValue(it->getOperand(1));
3810 for (unsigned Part = 0; Part < UF; ++Part) {
3813 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3814 cast<FCmpInst>(C)->copyFastMathFlags(&*it);
3816 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3821 propagateMetadata(Entry, &*it);
3825 case Instruction::Store:
3826 case Instruction::Load:
3827 vectorizeMemoryInstruction(&*it);
3829 case Instruction::ZExt:
3830 case Instruction::SExt:
3831 case Instruction::FPToUI:
3832 case Instruction::FPToSI:
3833 case Instruction::FPExt:
3834 case Instruction::PtrToInt:
3835 case Instruction::IntToPtr:
3836 case Instruction::SIToFP:
3837 case Instruction::UIToFP:
3838 case Instruction::Trunc:
3839 case Instruction::FPTrunc:
3840 case Instruction::BitCast: {
3841 CastInst *CI = dyn_cast<CastInst>(it);
3842 setDebugLocFromInst(Builder, &*it);
3843 /// Optimize the special case where the source is the induction
3844 /// variable. Notice that we can only optimize the 'trunc' case
3845 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3846 /// c. other casts depend on pointer size.
3847 if (CI->getOperand(0) == OldInduction &&
3848 it->getOpcode() == Instruction::Trunc) {
3849 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3851 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3852 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3854 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3855 for (unsigned Part = 0; Part < UF; ++Part)
3856 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3857 propagateMetadata(Entry, &*it);
3860 /// Vectorize casts.
3861 Type *DestTy = (VF == 1) ? CI->getType() :
3862 VectorType::get(CI->getType(), VF);
3864 VectorParts &A = getVectorValue(it->getOperand(0));
3865 for (unsigned Part = 0; Part < UF; ++Part)
3866 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3867 propagateMetadata(Entry, &*it);
3871 case Instruction::Call: {
3872 // Ignore dbg intrinsics.
3873 if (isa<DbgInfoIntrinsic>(it))
3875 setDebugLocFromInst(Builder, &*it);
3877 Module *M = BB->getParent()->getParent();
3878 CallInst *CI = cast<CallInst>(it);
3880 StringRef FnName = CI->getCalledFunction()->getName();
3881 Function *F = CI->getCalledFunction();
3882 Type *RetTy = ToVectorTy(CI->getType(), VF);
3883 SmallVector<Type *, 4> Tys;
3884 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3885 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3887 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3889 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3890 ID == Intrinsic::lifetime_start)) {
3891 scalarizeInstruction(&*it);
3894 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3895 // version of the instruction.
3896 // Is it beneficial to perform intrinsic call compared to lib call?
3897 bool NeedToScalarize;
3898 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3899 bool UseVectorIntrinsic =
3900 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3901 if (!UseVectorIntrinsic && NeedToScalarize) {
3902 scalarizeInstruction(&*it);
3906 for (unsigned Part = 0; Part < UF; ++Part) {
3907 SmallVector<Value *, 4> Args;
3908 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3909 Value *Arg = CI->getArgOperand(i);
3910 // Some intrinsics have a scalar argument - don't replace it with a
3912 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3913 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3914 Arg = VectorArg[Part];
3916 Args.push_back(Arg);
3920 if (UseVectorIntrinsic) {
3921 // Use vector version of the intrinsic.
3922 Type *TysForDecl[] = {CI->getType()};
3924 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3925 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3927 // Use vector version of the library call.
3928 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3929 assert(!VFnName.empty() && "Vector function name is empty.");
3930 VectorF = M->getFunction(VFnName);
3932 // Generate a declaration
3933 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3935 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3936 VectorF->copyAttributesFrom(F);
3939 assert(VectorF && "Can't create vector function.");
3940 Entry[Part] = Builder.CreateCall(VectorF, Args);
3943 propagateMetadata(Entry, &*it);
3948 // All other instructions are unsupported. Scalarize them.
3949 scalarizeInstruction(&*it);
3952 }// end of for_each instr.
3955 void InnerLoopVectorizer::updateAnalysis() {
3956 // Forget the original basic block.
3957 SE->forgetLoop(OrigLoop);
3959 // Update the dominator tree information.
3960 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3961 "Entry does not dominate exit.");
3963 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3964 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3965 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3967 // We don't predicate stores by this point, so the vector body should be a
3969 assert(LoopVectorBody.size() == 1 && "Expected single block loop!");
3970 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3972 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody.back());
3973 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3974 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3975 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3977 DEBUG(DT->verifyDomTree());
3980 /// \brief Check whether it is safe to if-convert this phi node.
3982 /// Phi nodes with constant expressions that can trap are not safe to if
3984 static bool canIfConvertPHINodes(BasicBlock *BB) {
3985 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3986 PHINode *Phi = dyn_cast<PHINode>(I);
3989 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3990 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3997 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3998 if (!EnableIfConversion) {
3999 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
4003 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
4005 // A list of pointers that we can safely read and write to.
4006 SmallPtrSet<Value *, 8> SafePointes;
4008 // Collect safe addresses.
4009 for (Loop::block_iterator BI = TheLoop->block_begin(),
4010 BE = TheLoop->block_end(); BI != BE; ++BI) {
4011 BasicBlock *BB = *BI;
4013 if (blockNeedsPredication(BB))
4016 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
4017 if (LoadInst *LI = dyn_cast<LoadInst>(I))
4018 SafePointes.insert(LI->getPointerOperand());
4019 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
4020 SafePointes.insert(SI->getPointerOperand());
4024 // Collect the blocks that need predication.
4025 BasicBlock *Header = TheLoop->getHeader();
4026 for (Loop::block_iterator BI = TheLoop->block_begin(),
4027 BE = TheLoop->block_end(); BI != BE; ++BI) {
4028 BasicBlock *BB = *BI;
4030 // We don't support switch statements inside loops.
4031 if (!isa<BranchInst>(BB->getTerminator())) {
4032 emitAnalysis(VectorizationReport(BB->getTerminator())
4033 << "loop contains a switch statement");
4037 // We must be able to predicate all blocks that need to be predicated.
4038 if (blockNeedsPredication(BB)) {
4039 if (!blockCanBePredicated(BB, SafePointes)) {
4040 emitAnalysis(VectorizationReport(BB->getTerminator())
4041 << "control flow cannot be substituted for a select");
4044 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
4045 emitAnalysis(VectorizationReport(BB->getTerminator())
4046 << "control flow cannot be substituted for a select");
4051 // We can if-convert this loop.
4055 bool LoopVectorizationLegality::canVectorize() {
4056 // We must have a loop in canonical form. Loops with indirectbr in them cannot
4057 // be canonicalized.
4058 if (!TheLoop->getLoopPreheader()) {
4060 VectorizationReport() <<
4061 "loop control flow is not understood by vectorizer");
4065 // We can only vectorize innermost loops.
4066 if (!TheLoop->empty()) {
4067 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
4071 // We must have a single backedge.
4072 if (TheLoop->getNumBackEdges() != 1) {
4074 VectorizationReport() <<
4075 "loop control flow is not understood by vectorizer");
4079 // We must have a single exiting block.
4080 if (!TheLoop->getExitingBlock()) {
4082 VectorizationReport() <<
4083 "loop control flow is not understood by vectorizer");
4087 // We only handle bottom-tested loops, i.e. loop in which the condition is
4088 // checked at the end of each iteration. With that we can assume that all
4089 // instructions in the loop are executed the same number of times.
4090 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
4092 VectorizationReport() <<
4093 "loop control flow is not understood by vectorizer");
4097 // We need to have a loop header.
4098 DEBUG(dbgs() << "LV: Found a loop: " <<
4099 TheLoop->getHeader()->getName() << '\n');
4101 // Check if we can if-convert non-single-bb loops.
4102 unsigned NumBlocks = TheLoop->getNumBlocks();
4103 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4104 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4108 // ScalarEvolution needs to be able to find the exit count.
4109 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
4110 if (ExitCount == SE->getCouldNotCompute()) {
4111 emitAnalysis(VectorizationReport() <<
4112 "could not determine number of loop iterations");
4113 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4117 // Check if we can vectorize the instructions and CFG in this loop.
4118 if (!canVectorizeInstrs()) {
4119 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4123 // Go over each instruction and look at memory deps.
4124 if (!canVectorizeMemory()) {
4125 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4129 // Collect all of the variables that remain uniform after vectorization.
4130 collectLoopUniforms();
4132 DEBUG(dbgs() << "LV: We can vectorize this loop"
4133 << (LAI->getRuntimePointerChecking()->Need
4134 ? " (with a runtime bound check)"
4138 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4140 // If an override option has been passed in for interleaved accesses, use it.
4141 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4142 UseInterleaved = EnableInterleavedMemAccesses;
4144 // Analyze interleaved memory accesses.
4146 InterleaveInfo.analyzeInterleaving(Strides);
4148 // Okay! We can vectorize. At this point we don't have any other mem analysis
4149 // which may limit our maximum vectorization factor, so just return true with
4154 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4155 if (Ty->isPointerTy())
4156 return DL.getIntPtrType(Ty);
4158 // It is possible that char's or short's overflow when we ask for the loop's
4159 // trip count, work around this by changing the type size.
4160 if (Ty->getScalarSizeInBits() < 32)
4161 return Type::getInt32Ty(Ty->getContext());
4166 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4167 Ty0 = convertPointerToIntegerType(DL, Ty0);
4168 Ty1 = convertPointerToIntegerType(DL, Ty1);
4169 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4174 /// \brief Check that the instruction has outside loop users and is not an
4175 /// identified reduction variable.
4176 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4177 SmallPtrSetImpl<Value *> &Reductions) {
4178 // Reduction instructions are allowed to have exit users. All other
4179 // instructions must not have external users.
4180 if (!Reductions.count(Inst))
4181 //Check that all of the users of the loop are inside the BB.
4182 for (User *U : Inst->users()) {
4183 Instruction *UI = cast<Instruction>(U);
4184 // This user may be a reduction exit value.
4185 if (!TheLoop->contains(UI)) {
4186 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4193 bool LoopVectorizationLegality::canVectorizeInstrs() {
4194 BasicBlock *Header = TheLoop->getHeader();
4196 // Look for the attribute signaling the absence of NaNs.
4197 Function &F = *Header->getParent();
4198 const DataLayout &DL = F.getParent()->getDataLayout();
4199 if (F.hasFnAttribute("no-nans-fp-math"))
4201 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4203 // For each block in the loop.
4204 for (Loop::block_iterator bb = TheLoop->block_begin(),
4205 be = TheLoop->block_end(); bb != be; ++bb) {
4207 // Scan the instructions in the block and look for hazards.
4208 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4211 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4212 Type *PhiTy = Phi->getType();
4213 // Check that this PHI type is allowed.
4214 if (!PhiTy->isIntegerTy() &&
4215 !PhiTy->isFloatingPointTy() &&
4216 !PhiTy->isPointerTy()) {
4217 emitAnalysis(VectorizationReport(&*it)
4218 << "loop control flow is not understood by vectorizer");
4219 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4223 // If this PHINode is not in the header block, then we know that we
4224 // can convert it to select during if-conversion. No need to check if
4225 // the PHIs in this block are induction or reduction variables.
4226 if (*bb != Header) {
4227 // Check that this instruction has no outside users or is an
4228 // identified reduction value with an outside user.
4229 if (!hasOutsideLoopUser(TheLoop, &*it, AllowedExit))
4231 emitAnalysis(VectorizationReport(&*it) <<
4232 "value could not be identified as "
4233 "an induction or reduction variable");
4237 // We only allow if-converted PHIs with exactly two incoming values.
4238 if (Phi->getNumIncomingValues() != 2) {
4239 emitAnalysis(VectorizationReport(&*it)
4240 << "control flow not understood by vectorizer");
4241 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4245 InductionDescriptor ID;
4246 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4247 Inductions[Phi] = ID;
4248 // Get the widest type.
4250 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4252 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4254 // Int inductions are special because we only allow one IV.
4255 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4256 ID.getStepValue()->isOne() &&
4257 isa<Constant>(ID.getStartValue()) &&
4258 cast<Constant>(ID.getStartValue())->isNullValue()) {
4259 // Use the phi node with the widest type as induction. Use the last
4260 // one if there are multiple (no good reason for doing this other
4261 // than it is expedient). We've checked that it begins at zero and
4262 // steps by one, so this is a canonical induction variable.
4263 if (!Induction || PhiTy == WidestIndTy)
4267 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4269 // Until we explicitly handle the case of an induction variable with
4270 // an outside loop user we have to give up vectorizing this loop.
4271 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4272 emitAnalysis(VectorizationReport(&*it) <<
4273 "use of induction value outside of the "
4274 "loop is not handled by vectorizer");
4281 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4283 if (Reductions[Phi].hasUnsafeAlgebra())
4284 Requirements->addUnsafeAlgebraInst(
4285 Reductions[Phi].getUnsafeAlgebraInst());
4286 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4290 emitAnalysis(VectorizationReport(&*it) <<
4291 "value that could not be identified as "
4292 "reduction is used outside the loop");
4293 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4295 }// end of PHI handling
4297 // We handle calls that:
4298 // * Are debug info intrinsics.
4299 // * Have a mapping to an IR intrinsic.
4300 // * Have a vector version available.
4301 CallInst *CI = dyn_cast<CallInst>(it);
4302 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4303 !(CI->getCalledFunction() && TLI &&
4304 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4305 emitAnalysis(VectorizationReport(&*it)
4306 << "call instruction cannot be vectorized");
4307 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4311 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4312 // second argument is the same (i.e. loop invariant)
4314 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4315 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4316 emitAnalysis(VectorizationReport(&*it)
4317 << "intrinsic instruction cannot be vectorized");
4318 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4323 // Check that the instruction return type is vectorizable.
4324 // Also, we can't vectorize extractelement instructions.
4325 if ((!VectorType::isValidElementType(it->getType()) &&
4326 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4327 emitAnalysis(VectorizationReport(&*it)
4328 << "instruction return type cannot be vectorized");
4329 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4333 // Check that the stored type is vectorizable.
4334 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4335 Type *T = ST->getValueOperand()->getType();
4336 if (!VectorType::isValidElementType(T)) {
4337 emitAnalysis(VectorizationReport(ST) <<
4338 "store instruction cannot be vectorized");
4341 if (EnableMemAccessVersioning)
4342 collectStridedAccess(ST);
4345 if (EnableMemAccessVersioning)
4346 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4347 collectStridedAccess(LI);
4349 // Reduction instructions are allowed to have exit users.
4350 // All other instructions must not have external users.
4351 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4352 emitAnalysis(VectorizationReport(&*it) <<
4353 "value cannot be used outside the loop");
4362 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4363 if (Inductions.empty()) {
4364 emitAnalysis(VectorizationReport()
4365 << "loop induction variable could not be identified");
4370 // Now we know the widest induction type, check if our found induction
4371 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4372 // will create another.
4373 if (Induction && WidestIndTy != Induction->getType())
4374 Induction = nullptr;
4379 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4380 Value *Ptr = nullptr;
4381 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4382 Ptr = LI->getPointerOperand();
4383 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4384 Ptr = SI->getPointerOperand();
4388 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4392 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4393 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4394 Strides[Ptr] = Stride;
4395 StrideSet.insert(Stride);
4398 void LoopVectorizationLegality::collectLoopUniforms() {
4399 // We now know that the loop is vectorizable!
4400 // Collect variables that will remain uniform after vectorization.
4401 std::vector<Value*> Worklist;
4402 BasicBlock *Latch = TheLoop->getLoopLatch();
4404 // Start with the conditional branch and walk up the block.
4405 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4407 // Also add all consecutive pointer values; these values will be uniform
4408 // after vectorization (and subsequent cleanup) and, until revectorization is
4409 // supported, all dependencies must also be uniform.
4410 for (Loop::block_iterator B = TheLoop->block_begin(),
4411 BE = TheLoop->block_end(); B != BE; ++B)
4412 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4414 if (I->getType()->isPointerTy() && isConsecutivePtr(&*I))
4415 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4417 while (!Worklist.empty()) {
4418 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4419 Worklist.pop_back();
4421 // Look at instructions inside this loop.
4422 // Stop when reaching PHI nodes.
4423 // TODO: we need to follow values all over the loop, not only in this block.
4424 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4427 // This is a known uniform.
4430 // Insert all operands.
4431 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4435 bool LoopVectorizationLegality::canVectorizeMemory() {
4436 LAI = &LAA->getInfo(TheLoop, Strides);
4437 auto &OptionalReport = LAI->getReport();
4439 emitAnalysis(VectorizationReport(*OptionalReport));
4440 if (!LAI->canVectorizeMemory())
4443 if (LAI->hasStoreToLoopInvariantAddress()) {
4445 VectorizationReport()
4446 << "write to a loop invariant address could not be vectorized");
4447 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4451 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4456 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4457 Value *In0 = const_cast<Value*>(V);
4458 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4462 return Inductions.count(PN);
4465 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4466 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4469 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4470 SmallPtrSetImpl<Value *> &SafePtrs) {
4472 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4473 // Check that we don't have a constant expression that can trap as operand.
4474 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4476 if (Constant *C = dyn_cast<Constant>(*OI))
4480 // We might be able to hoist the load.
4481 if (it->mayReadFromMemory()) {
4482 LoadInst *LI = dyn_cast<LoadInst>(it);
4485 if (!SafePtrs.count(LI->getPointerOperand())) {
4486 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4487 MaskedOp.insert(LI);
4494 // We don't predicate stores at the moment.
4495 if (it->mayWriteToMemory()) {
4496 StoreInst *SI = dyn_cast<StoreInst>(it);
4497 // We only support predication of stores in basic blocks with one
4502 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4503 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4505 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4506 !isSinglePredecessor) {
4507 // Build a masked store if it is legal for the target, otherwise scalarize
4509 bool isLegalMaskedOp =
4510 isLegalMaskedStore(SI->getValueOperand()->getType(),
4511 SI->getPointerOperand());
4512 if (isLegalMaskedOp) {
4514 MaskedOp.insert(SI);
4523 // The instructions below can trap.
4524 switch (it->getOpcode()) {
4526 case Instruction::UDiv:
4527 case Instruction::SDiv:
4528 case Instruction::URem:
4529 case Instruction::SRem:
4537 void InterleavedAccessInfo::collectConstStridedAccesses(
4538 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4539 const ValueToValueMap &Strides) {
4540 // Holds load/store instructions in program order.
4541 SmallVector<Instruction *, 16> AccessList;
4543 for (auto *BB : TheLoop->getBlocks()) {
4544 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4546 for (auto &I : *BB) {
4547 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4549 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4553 AccessList.push_back(&I);
4557 if (AccessList.empty())
4560 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4561 for (auto I : AccessList) {
4562 LoadInst *LI = dyn_cast<LoadInst>(I);
4563 StoreInst *SI = dyn_cast<StoreInst>(I);
4565 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4566 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4568 // The factor of the corresponding interleave group.
4569 unsigned Factor = std::abs(Stride);
4571 // Ignore the access if the factor is too small or too large.
4572 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4575 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4576 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4577 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4579 // An alignment of 0 means target ABI alignment.
4580 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4582 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4584 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4588 // Analyze interleaved accesses and collect them into interleave groups.
4590 // Notice that the vectorization on interleaved groups will change instruction
4591 // orders and may break dependences. But the memory dependence check guarantees
4592 // that there is no overlap between two pointers of different strides, element
4593 // sizes or underlying bases.
4595 // For pointers sharing the same stride, element size and underlying base, no
4596 // need to worry about Read-After-Write dependences and Write-After-Read
4599 // E.g. The RAW dependence: A[i] = a;
4601 // This won't exist as it is a store-load forwarding conflict, which has
4602 // already been checked and forbidden in the dependence check.
4604 // E.g. The WAR dependence: a = A[i]; // (1)
4606 // The store group of (2) is always inserted at or below (2), and the load group
4607 // of (1) is always inserted at or above (1). The dependence is safe.
4608 void InterleavedAccessInfo::analyzeInterleaving(
4609 const ValueToValueMap &Strides) {
4610 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4612 // Holds all the stride accesses.
4613 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4614 collectConstStridedAccesses(StrideAccesses, Strides);
4616 if (StrideAccesses.empty())
4619 // Holds all interleaved store groups temporarily.
4620 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4622 // Search the load-load/write-write pair B-A in bottom-up order and try to
4623 // insert B into the interleave group of A according to 3 rules:
4624 // 1. A and B have the same stride.
4625 // 2. A and B have the same memory object size.
4626 // 3. B belongs to the group according to the distance.
4628 // The bottom-up order can avoid breaking the Write-After-Write dependences
4629 // between two pointers of the same base.
4630 // E.g. A[i] = a; (1)
4633 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4634 // above (1), which guarantees that (1) is always above (2).
4635 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4637 Instruction *A = I->first;
4638 StrideDescriptor DesA = I->second;
4640 InterleaveGroup *Group = getInterleaveGroup(A);
4642 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4643 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4646 if (A->mayWriteToMemory())
4647 StoreGroups.insert(Group);
4649 for (auto II = std::next(I); II != E; ++II) {
4650 Instruction *B = II->first;
4651 StrideDescriptor DesB = II->second;
4653 // Ignore if B is already in a group or B is a different memory operation.
4654 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4657 // Check the rule 1 and 2.
4658 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4661 // Calculate the distance and prepare for the rule 3.
4662 const SCEVConstant *DistToA =
4663 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4667 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4669 // Skip if the distance is not multiple of size as they are not in the
4671 if (DistanceToA % static_cast<int>(DesA.Size))
4674 // The index of B is the index of A plus the related index to A.
4676 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4678 // Try to insert B into the group.
4679 if (Group->insertMember(B, IndexB, DesB.Align)) {
4680 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4681 << " into the interleave group with" << *A << '\n');
4682 InterleaveGroupMap[B] = Group;
4684 // Set the first load in program order as the insert position.
4685 if (B->mayReadFromMemory())
4686 Group->setInsertPos(B);
4688 } // Iteration on instruction B
4689 } // Iteration on instruction A
4691 // Remove interleaved store groups with gaps.
4692 for (InterleaveGroup *Group : StoreGroups)
4693 if (Group->getNumMembers() != Group->getFactor())
4694 releaseGroup(Group);
4697 LoopVectorizationCostModel::VectorizationFactor
4698 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4699 // Width 1 means no vectorize
4700 VectorizationFactor Factor = { 1U, 0U };
4701 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4702 emitAnalysis(VectorizationReport() <<
4703 "runtime pointer checks needed. Enable vectorization of this "
4704 "loop with '#pragma clang loop vectorize(enable)' when "
4705 "compiling with -Os/-Oz");
4707 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4711 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4712 emitAnalysis(VectorizationReport() <<
4713 "store that is conditionally executed prevents vectorization");
4714 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4718 // Find the trip count.
4719 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4720 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4722 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
4723 unsigned WidestType = getWidestType();
4724 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4725 unsigned MaxSafeDepDist = -1U;
4726 if (Legal->getMaxSafeDepDistBytes() != -1U)
4727 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4728 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4729 WidestRegister : MaxSafeDepDist);
4730 unsigned MaxVectorSize = WidestRegister / WidestType;
4731 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4732 DEBUG(dbgs() << "LV: The Widest register is: "
4733 << WidestRegister << " bits.\n");
4735 if (MaxVectorSize == 0) {
4736 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4740 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4741 " into one vector!");
4743 unsigned VF = MaxVectorSize;
4745 // If we optimize the program for size, avoid creating the tail loop.
4747 // If we are unable to calculate the trip count then don't try to vectorize.
4750 (VectorizationReport() <<
4751 "unable to calculate the loop count due to complex control flow");
4752 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4756 // Find the maximum SIMD width that can fit within the trip count.
4757 VF = TC % MaxVectorSize;
4762 // If the trip count that we found modulo the vectorization factor is not
4763 // zero then we require a tail.
4764 emitAnalysis(VectorizationReport() <<
4765 "cannot optimize for size and vectorize at the "
4766 "same time. Enable vectorization of this loop "
4767 "with '#pragma clang loop vectorize(enable)' "
4768 "when compiling with -Os/-Oz");
4769 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4774 int UserVF = Hints->getWidth();
4776 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4777 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4779 Factor.Width = UserVF;
4783 float Cost = expectedCost(1);
4785 const float ScalarCost = Cost;
4788 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4790 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4791 // Ignore scalar width, because the user explicitly wants vectorization.
4792 if (ForceVectorization && VF > 1) {
4794 Cost = expectedCost(Width) / (float)Width;
4797 for (unsigned i=2; i <= VF; i*=2) {
4798 // Notice that the vector loop needs to be executed less times, so
4799 // we need to divide the cost of the vector loops by the width of
4800 // the vector elements.
4801 float VectorCost = expectedCost(i) / (float)i;
4802 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4803 (int)VectorCost << ".\n");
4804 if (VectorCost < Cost) {
4810 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4811 << "LV: Vectorization seems to be not beneficial, "
4812 << "but was forced by a user.\n");
4813 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4814 Factor.Width = Width;
4815 Factor.Cost = Width * Cost;
4819 unsigned LoopVectorizationCostModel::getWidestType() {
4820 unsigned MaxWidth = 8;
4821 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4824 for (Loop::block_iterator bb = TheLoop->block_begin(),
4825 be = TheLoop->block_end(); bb != be; ++bb) {
4826 BasicBlock *BB = *bb;
4828 // For each instruction in the loop.
4829 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4830 Type *T = it->getType();
4832 // Skip ignored values.
4833 if (ValuesToIgnore.count(&*it))
4836 // Only examine Loads, Stores and PHINodes.
4837 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4840 // Examine PHI nodes that are reduction variables. Update the type to
4841 // account for the recurrence type.
4842 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4843 if (!Legal->getReductionVars()->count(PN))
4845 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4846 T = RdxDesc.getRecurrenceType();
4849 // Examine the stored values.
4850 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4851 T = ST->getValueOperand()->getType();
4853 // Ignore loaded pointer types and stored pointer types that are not
4854 // consecutive. However, we do want to take consecutive stores/loads of
4855 // pointer vectors into account.
4856 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&*it))
4859 MaxWidth = std::max(MaxWidth,
4860 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4867 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4869 unsigned LoopCost) {
4871 // -- The interleave heuristics --
4872 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4873 // There are many micro-architectural considerations that we can't predict
4874 // at this level. For example, frontend pressure (on decode or fetch) due to
4875 // code size, or the number and capabilities of the execution ports.
4877 // We use the following heuristics to select the interleave count:
4878 // 1. If the code has reductions, then we interleave to break the cross
4879 // iteration dependency.
4880 // 2. If the loop is really small, then we interleave to reduce the loop
4882 // 3. We don't interleave if we think that we will spill registers to memory
4883 // due to the increased register pressure.
4885 // When we optimize for size, we don't interleave.
4889 // We used the distance for the interleave count.
4890 if (Legal->getMaxSafeDepDistBytes() != -1U)
4893 // Do not interleave loops with a relatively small trip count.
4894 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4895 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4898 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4899 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4903 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4904 TargetNumRegisters = ForceTargetNumScalarRegs;
4906 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4907 TargetNumRegisters = ForceTargetNumVectorRegs;
4910 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4911 // We divide by these constants so assume that we have at least one
4912 // instruction that uses at least one register.
4913 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4914 R.NumInstructions = std::max(R.NumInstructions, 1U);
4916 // We calculate the interleave count using the following formula.
4917 // Subtract the number of loop invariants from the number of available
4918 // registers. These registers are used by all of the interleaved instances.
4919 // Next, divide the remaining registers by the number of registers that is
4920 // required by the loop, in order to estimate how many parallel instances
4921 // fit without causing spills. All of this is rounded down if necessary to be
4922 // a power of two. We want power of two interleave count to simplify any
4923 // addressing operations or alignment considerations.
4924 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4927 // Don't count the induction variable as interleaved.
4928 if (EnableIndVarRegisterHeur)
4929 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4930 std::max(1U, (R.MaxLocalUsers - 1)));
4932 // Clamp the interleave ranges to reasonable counts.
4933 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4935 // Check if the user has overridden the max.
4937 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4938 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4940 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4941 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4944 // If we did not calculate the cost for VF (because the user selected the VF)
4945 // then we calculate the cost of VF here.
4947 LoopCost = expectedCost(VF);
4949 // Clamp the calculated IC to be between the 1 and the max interleave count
4950 // that the target allows.
4951 if (IC > MaxInterleaveCount)
4952 IC = MaxInterleaveCount;
4956 // Interleave if we vectorized this loop and there is a reduction that could
4957 // benefit from interleaving.
4958 if (VF > 1 && Legal->getReductionVars()->size()) {
4959 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4963 // Note that if we've already vectorized the loop we will have done the
4964 // runtime check and so interleaving won't require further checks.
4965 bool InterleavingRequiresRuntimePointerCheck =
4966 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4968 // We want to interleave small loops in order to reduce the loop overhead and
4969 // potentially expose ILP opportunities.
4970 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4971 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4972 // We assume that the cost overhead is 1 and we use the cost model
4973 // to estimate the cost of the loop and interleave until the cost of the
4974 // loop overhead is about 5% of the cost of the loop.
4976 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4978 // Interleave until store/load ports (estimated by max interleave count) are
4980 unsigned NumStores = Legal->getNumStores();
4981 unsigned NumLoads = Legal->getNumLoads();
4982 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4983 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4985 // If we have a scalar reduction (vector reductions are already dealt with
4986 // by this point), we can increase the critical path length if the loop
4987 // we're interleaving is inside another loop. Limit, by default to 2, so the
4988 // critical path only gets increased by one reduction operation.
4989 if (Legal->getReductionVars()->size() &&
4990 TheLoop->getLoopDepth() > 1) {
4991 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4992 SmallIC = std::min(SmallIC, F);
4993 StoresIC = std::min(StoresIC, F);
4994 LoadsIC = std::min(LoadsIC, F);
4997 if (EnableLoadStoreRuntimeInterleave &&
4998 std::max(StoresIC, LoadsIC) > SmallIC) {
4999 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
5000 return std::max(StoresIC, LoadsIC);
5003 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
5007 // Interleave if this is a large loop (small loops are already dealt with by
5009 // point) that could benefit from interleaving.
5010 bool HasReductions = (Legal->getReductionVars()->size() > 0);
5011 if (TTI.enableAggressiveInterleaving(HasReductions)) {
5012 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
5016 DEBUG(dbgs() << "LV: Not Interleaving.\n");
5020 LoopVectorizationCostModel::RegisterUsage
5021 LoopVectorizationCostModel::calculateRegisterUsage() {
5022 // This function calculates the register usage by measuring the highest number
5023 // of values that are alive at a single location. Obviously, this is a very
5024 // rough estimation. We scan the loop in a topological order in order and
5025 // assign a number to each instruction. We use RPO to ensure that defs are
5026 // met before their users. We assume that each instruction that has in-loop
5027 // users starts an interval. We record every time that an in-loop value is
5028 // used, so we have a list of the first and last occurrences of each
5029 // instruction. Next, we transpose this data structure into a multi map that
5030 // holds the list of intervals that *end* at a specific location. This multi
5031 // map allows us to perform a linear search. We scan the instructions linearly
5032 // and record each time that a new interval starts, by placing it in a set.
5033 // If we find this value in the multi-map then we remove it from the set.
5034 // The max register usage is the maximum size of the set.
5035 // We also search for instructions that are defined outside the loop, but are
5036 // used inside the loop. We need this number separately from the max-interval
5037 // usage number because when we unroll, loop-invariant values do not take
5039 LoopBlocksDFS DFS(TheLoop);
5043 R.NumInstructions = 0;
5045 // Each 'key' in the map opens a new interval. The values
5046 // of the map are the index of the 'last seen' usage of the
5047 // instruction that is the key.
5048 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5049 // Maps instruction to its index.
5050 DenseMap<unsigned, Instruction*> IdxToInstr;
5051 // Marks the end of each interval.
5052 IntervalMap EndPoint;
5053 // Saves the list of instruction indices that are used in the loop.
5054 SmallSet<Instruction*, 8> Ends;
5055 // Saves the list of values that are used in the loop but are
5056 // defined outside the loop, such as arguments and constants.
5057 SmallPtrSet<Value*, 8> LoopInvariants;
5060 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5061 be = DFS.endRPO(); bb != be; ++bb) {
5062 R.NumInstructions += (*bb)->size();
5063 for (Instruction &I : **bb) {
5064 IdxToInstr[Index++] = &I;
5066 // Save the end location of each USE.
5067 for (unsigned i = 0; i < I.getNumOperands(); ++i) {
5068 Value *U = I.getOperand(i);
5069 Instruction *Instr = dyn_cast<Instruction>(U);
5071 // Ignore non-instruction values such as arguments, constants, etc.
5072 if (!Instr) continue;
5074 // If this instruction is outside the loop then record it and continue.
5075 if (!TheLoop->contains(Instr)) {
5076 LoopInvariants.insert(Instr);
5080 // Overwrite previous end points.
5081 EndPoint[Instr] = Index;
5087 // Saves the list of intervals that end with the index in 'key'.
5088 typedef SmallVector<Instruction*, 2> InstrList;
5089 DenseMap<unsigned, InstrList> TransposeEnds;
5091 // Transpose the EndPoints to a list of values that end at each index.
5092 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5094 TransposeEnds[it->second].push_back(it->first);
5096 SmallSet<Instruction*, 8> OpenIntervals;
5097 unsigned MaxUsage = 0;
5100 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5101 for (unsigned int i = 0; i < Index; ++i) {
5102 Instruction *I = IdxToInstr[i];
5103 // Ignore instructions that are never used within the loop.
5104 if (!Ends.count(I)) continue;
5106 // Skip ignored values.
5107 if (ValuesToIgnore.count(I))
5110 // Remove all of the instructions that end at this location.
5111 InstrList &List = TransposeEnds[i];
5112 for (unsigned int j=0, e = List.size(); j < e; ++j)
5113 OpenIntervals.erase(List[j]);
5115 // Count the number of live interals.
5116 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5118 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5119 OpenIntervals.size() << '\n');
5121 // Add the current instruction to the list of open intervals.
5122 OpenIntervals.insert(I);
5125 unsigned Invariant = LoopInvariants.size();
5126 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5127 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5128 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5130 R.LoopInvariantRegs = Invariant;
5131 R.MaxLocalUsers = MaxUsage;
5135 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5139 for (Loop::block_iterator bb = TheLoop->block_begin(),
5140 be = TheLoop->block_end(); bb != be; ++bb) {
5141 unsigned BlockCost = 0;
5142 BasicBlock *BB = *bb;
5144 // For each instruction in the old loop.
5145 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5146 // Skip dbg intrinsics.
5147 if (isa<DbgInfoIntrinsic>(it))
5150 // Skip ignored values.
5151 if (ValuesToIgnore.count(&*it))
5154 unsigned C = getInstructionCost(&*it, VF);
5156 // Check if we should override the cost.
5157 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5158 C = ForceTargetInstructionCost;
5161 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5162 VF << " For instruction: " << *it << '\n');
5165 // We assume that if-converted blocks have a 50% chance of being executed.
5166 // When the code is scalar then some of the blocks are avoided due to CF.
5167 // When the code is vectorized we execute all code paths.
5168 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5177 /// \brief Check whether the address computation for a non-consecutive memory
5178 /// access looks like an unlikely candidate for being merged into the indexing
5181 /// We look for a GEP which has one index that is an induction variable and all
5182 /// other indices are loop invariant. If the stride of this access is also
5183 /// within a small bound we decide that this address computation can likely be
5184 /// merged into the addressing mode.
5185 /// In all other cases, we identify the address computation as complex.
5186 static bool isLikelyComplexAddressComputation(Value *Ptr,
5187 LoopVectorizationLegality *Legal,
5188 ScalarEvolution *SE,
5189 const Loop *TheLoop) {
5190 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5194 // We are looking for a gep with all loop invariant indices except for one
5195 // which should be an induction variable.
5196 unsigned NumOperands = Gep->getNumOperands();
5197 for (unsigned i = 1; i < NumOperands; ++i) {
5198 Value *Opd = Gep->getOperand(i);
5199 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5200 !Legal->isInductionVariable(Opd))
5204 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5205 // can likely be merged into the address computation.
5206 unsigned MaxMergeDistance = 64;
5208 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5212 // Check the step is constant.
5213 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5214 // Calculate the pointer stride and check if it is consecutive.
5215 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5219 const APInt &APStepVal = C->getValue()->getValue();
5221 // Huge step value - give up.
5222 if (APStepVal.getBitWidth() > 64)
5225 int64_t StepVal = APStepVal.getSExtValue();
5227 return StepVal > MaxMergeDistance;
5230 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5231 return Legal->hasStride(I->getOperand(0)) ||
5232 Legal->hasStride(I->getOperand(1));
5236 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5237 // If we know that this instruction will remain uniform, check the cost of
5238 // the scalar version.
5239 if (Legal->isUniformAfterVectorization(I))
5242 Type *RetTy = I->getType();
5243 if (VF > 1 && MinBWs.count(I))
5244 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
5245 Type *VectorTy = ToVectorTy(RetTy, VF);
5247 // TODO: We need to estimate the cost of intrinsic calls.
5248 switch (I->getOpcode()) {
5249 case Instruction::GetElementPtr:
5250 // We mark this instruction as zero-cost because the cost of GEPs in
5251 // vectorized code depends on whether the corresponding memory instruction
5252 // is scalarized or not. Therefore, we handle GEPs with the memory
5253 // instruction cost.
5255 case Instruction::Br: {
5256 return TTI.getCFInstrCost(I->getOpcode());
5258 case Instruction::PHI:
5259 //TODO: IF-converted IFs become selects.
5261 case Instruction::Add:
5262 case Instruction::FAdd:
5263 case Instruction::Sub:
5264 case Instruction::FSub:
5265 case Instruction::Mul:
5266 case Instruction::FMul:
5267 case Instruction::UDiv:
5268 case Instruction::SDiv:
5269 case Instruction::FDiv:
5270 case Instruction::URem:
5271 case Instruction::SRem:
5272 case Instruction::FRem:
5273 case Instruction::Shl:
5274 case Instruction::LShr:
5275 case Instruction::AShr:
5276 case Instruction::And:
5277 case Instruction::Or:
5278 case Instruction::Xor: {
5279 // Since we will replace the stride by 1 the multiplication should go away.
5280 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5282 // Certain instructions can be cheaper to vectorize if they have a constant
5283 // second vector operand. One example of this are shifts on x86.
5284 TargetTransformInfo::OperandValueKind Op1VK =
5285 TargetTransformInfo::OK_AnyValue;
5286 TargetTransformInfo::OperandValueKind Op2VK =
5287 TargetTransformInfo::OK_AnyValue;
5288 TargetTransformInfo::OperandValueProperties Op1VP =
5289 TargetTransformInfo::OP_None;
5290 TargetTransformInfo::OperandValueProperties Op2VP =
5291 TargetTransformInfo::OP_None;
5292 Value *Op2 = I->getOperand(1);
5294 // Check for a splat of a constant or for a non uniform vector of constants.
5295 if (isa<ConstantInt>(Op2)) {
5296 ConstantInt *CInt = cast<ConstantInt>(Op2);
5297 if (CInt && CInt->getValue().isPowerOf2())
5298 Op2VP = TargetTransformInfo::OP_PowerOf2;
5299 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5300 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5301 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5302 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5304 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5305 if (CInt && CInt->getValue().isPowerOf2())
5306 Op2VP = TargetTransformInfo::OP_PowerOf2;
5307 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5311 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5314 case Instruction::Select: {
5315 SelectInst *SI = cast<SelectInst>(I);
5316 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5317 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5318 Type *CondTy = SI->getCondition()->getType();
5320 CondTy = VectorType::get(CondTy, VF);
5322 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5324 case Instruction::ICmp:
5325 case Instruction::FCmp: {
5326 Type *ValTy = I->getOperand(0)->getType();
5327 if (VF > 1 && MinBWs.count(dyn_cast<Instruction>(I->getOperand(0))))
5328 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[I]);
5329 VectorTy = ToVectorTy(ValTy, VF);
5330 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5332 case Instruction::Store:
5333 case Instruction::Load: {
5334 StoreInst *SI = dyn_cast<StoreInst>(I);
5335 LoadInst *LI = dyn_cast<LoadInst>(I);
5336 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5338 VectorTy = ToVectorTy(ValTy, VF);
5340 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5341 unsigned AS = SI ? SI->getPointerAddressSpace() :
5342 LI->getPointerAddressSpace();
5343 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5344 // We add the cost of address computation here instead of with the gep
5345 // instruction because only here we know whether the operation is
5348 return TTI.getAddressComputationCost(VectorTy) +
5349 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5351 // For an interleaved access, calculate the total cost of the whole
5352 // interleave group.
5353 if (Legal->isAccessInterleaved(I)) {
5354 auto Group = Legal->getInterleavedAccessGroup(I);
5355 assert(Group && "Fail to get an interleaved access group.");
5357 // Only calculate the cost once at the insert position.
5358 if (Group->getInsertPos() != I)
5361 unsigned InterleaveFactor = Group->getFactor();
5363 VectorType::get(VectorTy->getVectorElementType(),
5364 VectorTy->getVectorNumElements() * InterleaveFactor);
5366 // Holds the indices of existing members in an interleaved load group.
5367 // An interleaved store group doesn't need this as it dones't allow gaps.
5368 SmallVector<unsigned, 4> Indices;
5370 for (unsigned i = 0; i < InterleaveFactor; i++)
5371 if (Group->getMember(i))
5372 Indices.push_back(i);
5375 // Calculate the cost of the whole interleaved group.
5376 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5377 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5378 Group->getAlignment(), AS);
5380 if (Group->isReverse())
5382 Group->getNumMembers() *
5383 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5385 // FIXME: The interleaved load group with a huge gap could be even more
5386 // expensive than scalar operations. Then we could ignore such group and
5387 // use scalar operations instead.
5391 // Scalarized loads/stores.
5392 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5393 bool Reverse = ConsecutiveStride < 0;
5394 const DataLayout &DL = I->getModule()->getDataLayout();
5395 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5396 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5397 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5398 bool IsComplexComputation =
5399 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5401 // The cost of extracting from the value vector and pointer vector.
5402 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5403 for (unsigned i = 0; i < VF; ++i) {
5404 // The cost of extracting the pointer operand.
5405 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5406 // In case of STORE, the cost of ExtractElement from the vector.
5407 // In case of LOAD, the cost of InsertElement into the returned
5409 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5410 Instruction::InsertElement,
5414 // The cost of the scalar loads/stores.
5415 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5416 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5421 // Wide load/stores.
5422 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5423 if (Legal->isMaskRequired(I))
5424 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5427 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5430 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5434 case Instruction::ZExt:
5435 case Instruction::SExt:
5436 case Instruction::FPToUI:
5437 case Instruction::FPToSI:
5438 case Instruction::FPExt:
5439 case Instruction::PtrToInt:
5440 case Instruction::IntToPtr:
5441 case Instruction::SIToFP:
5442 case Instruction::UIToFP:
5443 case Instruction::Trunc:
5444 case Instruction::FPTrunc:
5445 case Instruction::BitCast: {
5446 // We optimize the truncation of induction variable.
5447 // The cost of these is the same as the scalar operation.
5448 if (I->getOpcode() == Instruction::Trunc &&
5449 Legal->isInductionVariable(I->getOperand(0)))
5450 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5451 I->getOperand(0)->getType());
5453 Type *SrcScalarTy = I->getOperand(0)->getType();
5454 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
5455 if (VF > 1 && MinBWs.count(I)) {
5456 // This cast is going to be shrunk. This may remove the cast or it might
5457 // turn it into slightly different cast. For example, if MinBW == 16,
5458 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
5460 // Calculate the modified src and dest types.
5461 Type *MinVecTy = VectorTy;
5462 if (I->getOpcode() == Instruction::Trunc) {
5463 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
5464 VectorTy = largestIntegerVectorType(ToVectorTy(I->getType(), VF),
5466 } else if (I->getOpcode() == Instruction::ZExt ||
5467 I->getOpcode() == Instruction::SExt) {
5468 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
5469 VectorTy = smallestIntegerVectorType(ToVectorTy(I->getType(), VF),
5474 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5476 case Instruction::Call: {
5477 bool NeedToScalarize;
5478 CallInst *CI = cast<CallInst>(I);
5479 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5480 if (getIntrinsicIDForCall(CI, TLI))
5481 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5485 // We are scalarizing the instruction. Return the cost of the scalar
5486 // instruction, plus the cost of insert and extract into vector
5487 // elements, times the vector width.
5490 if (!RetTy->isVoidTy() && VF != 1) {
5491 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5493 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5496 // The cost of inserting the results plus extracting each one of the
5498 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5501 // The cost of executing VF copies of the scalar instruction. This opcode
5502 // is unknown. Assume that it is the same as 'mul'.
5503 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5509 char LoopVectorize::ID = 0;
5510 static const char lv_name[] = "Loop Vectorization";
5511 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5512 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5513 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
5514 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
5515 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
5516 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5517 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5518 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5519 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5520 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5521 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5522 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5523 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5524 INITIALIZE_PASS_DEPENDENCY(DemandedBits)
5525 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5528 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5529 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5533 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5534 // Check for a store.
5535 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5536 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5538 // Check for a load.
5539 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5540 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5546 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5547 bool IfPredicateStore) {
5548 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5549 // Holds vector parameters or scalars, in case of uniform vals.
5550 SmallVector<VectorParts, 4> Params;
5552 setDebugLocFromInst(Builder, Instr);
5554 // Find all of the vectorized parameters.
5555 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5556 Value *SrcOp = Instr->getOperand(op);
5558 // If we are accessing the old induction variable, use the new one.
5559 if (SrcOp == OldInduction) {
5560 Params.push_back(getVectorValue(SrcOp));
5564 // Try using previously calculated values.
5565 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5567 // If the src is an instruction that appeared earlier in the basic block
5568 // then it should already be vectorized.
5569 if (SrcInst && OrigLoop->contains(SrcInst)) {
5570 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5571 // The parameter is a vector value from earlier.
5572 Params.push_back(WidenMap.get(SrcInst));
5574 // The parameter is a scalar from outside the loop. Maybe even a constant.
5575 VectorParts Scalars;
5576 Scalars.append(UF, SrcOp);
5577 Params.push_back(Scalars);
5581 assert(Params.size() == Instr->getNumOperands() &&
5582 "Invalid number of operands");
5584 // Does this instruction return a value ?
5585 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5587 Value *UndefVec = IsVoidRetTy ? nullptr :
5588 UndefValue::get(Instr->getType());
5589 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5590 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5593 if (IfPredicateStore) {
5594 assert(Instr->getParent()->getSinglePredecessor() &&
5595 "Only support single predecessor blocks");
5596 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5597 Instr->getParent());
5600 // For each vector unroll 'part':
5601 for (unsigned Part = 0; Part < UF; ++Part) {
5602 // For each scalar that we create:
5604 // Start an "if (pred) a[i] = ..." block.
5605 Value *Cmp = nullptr;
5606 if (IfPredicateStore) {
5607 if (Cond[Part]->getType()->isVectorTy())
5609 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5610 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5611 ConstantInt::get(Cond[Part]->getType(), 1));
5614 Instruction *Cloned = Instr->clone();
5616 Cloned->setName(Instr->getName() + ".cloned");
5617 // Replace the operands of the cloned instructions with extracted scalars.
5618 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5619 Value *Op = Params[op][Part];
5620 Cloned->setOperand(op, Op);
5623 // Place the cloned scalar in the new loop.
5624 Builder.Insert(Cloned);
5626 // If the original scalar returns a value we need to place it in a vector
5627 // so that future users will be able to use it.
5629 VecResults[Part] = Cloned;
5632 if (IfPredicateStore)
5633 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
5638 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5639 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5640 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5642 return scalarizeInstruction(Instr, IfPredicateStore);
5645 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5649 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5653 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5654 // When unrolling and the VF is 1, we only need to add a simple scalar.
5655 Type *ITy = Val->getType();
5656 assert(!ITy->isVectorTy() && "Val must be a scalar");
5657 Constant *C = ConstantInt::get(ITy, StartIdx);
5658 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");