Bug Summary

File:lib/Transforms/Vectorize/LoopVectorize.cpp
Location:line 4869, column 75
Description:Division by zero

Annotated Source Code

1//===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
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.
14//
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.
18//
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.
28//
29//===----------------------------------------------------------------------===//
30//
31// The reduction-variable vectorization is based on the paper:
32// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33//
34// Variable uniformity checks are inspired by:
35// Karrenberg, R. and Hack, S. Whole Function Vectorization.
36//
37// The interleaved access vectorization is based on the paper:
38// Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
39// Data for SIMD
40//
41// Other ideas/concepts are from:
42// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
43//
44// S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45// Vectorizing Compilers.
46//
47//===----------------------------------------------------------------------===//
48
49#include "llvm/Transforms/Vectorize.h"
50#include "llvm/ADT/DenseMap.h"
51#include "llvm/ADT/EquivalenceClasses.h"
52#include "llvm/ADT/Hashing.h"
53#include "llvm/ADT/MapVector.h"
54#include "llvm/ADT/SetVector.h"
55#include "llvm/ADT/SmallPtrSet.h"
56#include "llvm/ADT/SmallSet.h"
57#include "llvm/ADT/SmallVector.h"
58#include "llvm/ADT/Statistic.h"
59#include "llvm/ADT/StringExtras.h"
60#include "llvm/Analysis/AliasAnalysis.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/LoopAccessAnalysis.h"
66#include "llvm/Analysis/LoopInfo.h"
67#include "llvm/Analysis/LoopIterator.h"
68#include "llvm/Analysis/LoopPass.h"
69#include "llvm/Analysis/ScalarEvolution.h"
70#include "llvm/Analysis/ScalarEvolutionExpander.h"
71#include "llvm/Analysis/ScalarEvolutionExpressions.h"
72#include "llvm/Analysis/TargetTransformInfo.h"
73#include "llvm/Analysis/ValueTracking.h"
74#include "llvm/IR/Constants.h"
75#include "llvm/IR/DataLayout.h"
76#include "llvm/IR/DebugInfo.h"
77#include "llvm/IR/DerivedTypes.h"
78#include "llvm/IR/DiagnosticInfo.h"
79#include "llvm/IR/Dominators.h"
80#include "llvm/IR/Function.h"
81#include "llvm/IR/IRBuilder.h"
82#include "llvm/IR/Instructions.h"
83#include "llvm/IR/IntrinsicInst.h"
84#include "llvm/IR/LLVMContext.h"
85#include "llvm/IR/Module.h"
86#include "llvm/IR/PatternMatch.h"
87#include "llvm/IR/Type.h"
88#include "llvm/IR/Value.h"
89#include "llvm/IR/ValueHandle.h"
90#include "llvm/IR/Verifier.h"
91#include "llvm/Pass.h"
92#include "llvm/Support/BranchProbability.h"
93#include "llvm/Support/CommandLine.h"
94#include "llvm/Support/Debug.h"
95#include "llvm/Support/raw_ostream.h"
96#include "llvm/Transforms/Scalar.h"
97#include "llvm/Transforms/Utils/BasicBlockUtils.h"
98#include "llvm/Transforms/Utils/Local.h"
99#include "llvm/Analysis/VectorUtils.h"
100#include "llvm/Transforms/Utils/LoopUtils.h"
101#include <algorithm>
102#include <map>
103#include <tuple>
104
105using namespace llvm;
106using namespace llvm::PatternMatch;
107
108#define LV_NAME"loop-vectorize" "loop-vectorize"
109#define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize"
110
111STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = { "loop-vectorize", "Number of loops vectorized"
, 0, 0 };
112STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = { "loop-vectorize", "Number of loops analyzed for vectorization"
, 0, 0 };
113
114static cl::opt<bool>
115EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
117
118/// We don't vectorize loops with a known constant trip count below this number.
119static cl::opt<unsigned>
120TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
121 cl::Hidden,
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
124 "value."));
125
126/// This enables versioning on the strides of symbolically striding memory
127/// accesses in code like the following.
128/// for (i = 0; i < N; ++i)
129/// A[i * Stride1] += B[i * Stride2] ...
130///
131/// Will be roughly translated to
132/// if (Stride1 == 1 && Stride2 == 1) {
133/// for (i = 0; i < N; i+=4)
134/// A[i:i+3] += ...
135/// } else
136/// ...
137static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
140
141static cl::opt<bool> EnableInterleavedMemAccesses(
142 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
143 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
144
145/// Maximum factor for an interleaved memory access.
146static cl::opt<unsigned> MaxInterleaveGroupFactor(
147 "max-interleave-group-factor", cl::Hidden,
148 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
149 cl::init(8));
150
151/// We don't unroll loops with a known constant trip count below this number.
152static const unsigned TinyTripCountUnrollThreshold = 128;
153
154static cl::opt<unsigned> ForceTargetNumScalarRegs(
155 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's number of scalar registers."));
157
158static cl::opt<unsigned> ForceTargetNumVectorRegs(
159 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of vector registers."));
161
162/// Maximum vectorization interleave count.
163static const unsigned MaxInterleaveFactor = 16;
164
165static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
166 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's max interleave factor for "
168 "scalar loops."));
169
170static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
171 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
172 cl::desc("A flag that overrides the target's max interleave factor for "
173 "vectorized loops."));
174
175static cl::opt<unsigned> ForceTargetInstructionCost(
176 "force-target-instruction-cost", cl::init(0), cl::Hidden,
177 cl::desc("A flag that overrides the target's expected cost for "
178 "an instruction to a single constant value. Mostly "
179 "useful for getting consistent testing."));
180
181static cl::opt<unsigned> SmallLoopCost(
182 "small-loop-cost", cl::init(20), cl::Hidden,
183 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
184
185static cl::opt<bool> LoopVectorizeWithBlockFrequency(
186 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
187 cl::desc("Enable the use of the block frequency analysis to access PGO "
188 "heuristics minimizing code growth in cold regions and being more "
189 "aggressive in hot regions."));
190
191// Runtime unroll loops for load/store throughput.
192static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
193 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
194 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
195
196/// The number of stores in a loop that are allowed to need predication.
197static cl::opt<unsigned> NumberOfStoresToPredicate(
198 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
199 cl::desc("Max number of stores to be predicated behind an if."));
200
201static cl::opt<bool> EnableIndVarRegisterHeur(
202 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
203 cl::desc("Count the induction variable only once when unrolling"));
204
205static cl::opt<bool> EnableCondStoresVectorization(
206 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
207 cl::desc("Enable if predication of stores during vectorization."));
208
209static cl::opt<unsigned> MaxNestedScalarReductionUF(
210 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
211 cl::desc("The maximum unroll factor to use when unrolling a scalar "
212 "reduction in a nested loop."));
213
214namespace {
215
216// Forward declarations.
217class LoopVectorizationLegality;
218class LoopVectorizationCostModel;
219class LoopVectorizeHints;
220
221/// \brief This modifies LoopAccessReport to initialize message with
222/// loop-vectorizer-specific part.
223class VectorizationReport : public LoopAccessReport {
224public:
225 VectorizationReport(Instruction *I = nullptr)
226 : LoopAccessReport("loop not vectorized: ", I) {}
227
228 /// \brief This allows promotion of the loop-access analysis report into the
229 /// loop-vectorizer report. It modifies the message to add the
230 /// loop-vectorizer-specific part of the message.
231 explicit VectorizationReport(const LoopAccessReport &R)
232 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
233 R.getInstr()) {}
234};
235
236/// A helper function for converting Scalar types to vector types.
237/// If the incoming type is void, we return void. If the VF is 1, we return
238/// the scalar type.
239static Type* ToVectorTy(Type *Scalar, unsigned VF) {
240 if (Scalar->isVoidTy() || VF == 1)
241 return Scalar;
242 return VectorType::get(Scalar, VF);
243}
244
245/// InnerLoopVectorizer vectorizes loops which contain only one basic
246/// block to a specified vectorization factor (VF).
247/// This class performs the widening of scalars into vectors, or multiple
248/// scalars. This class also implements the following features:
249/// * It inserts an epilogue loop for handling loops that don't have iteration
250/// counts that are known to be a multiple of the vectorization factor.
251/// * It handles the code generation for reduction variables.
252/// * Scalarization (implementation using scalars) of un-vectorizable
253/// instructions.
254/// InnerLoopVectorizer does not perform any vectorization-legality
255/// checks, and relies on the caller to check for the different legality
256/// aspects. The InnerLoopVectorizer relies on the
257/// LoopVectorizationLegality class to provide information about the induction
258/// and reduction variables that were found to a given vectorization factor.
259class InnerLoopVectorizer {
260public:
261 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
262 DominatorTree *DT, const TargetLibraryInfo *TLI,
263 const TargetTransformInfo *TTI, unsigned VecWidth,
264 unsigned UnrollFactor)
265 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
266 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
267 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
268 Legal(nullptr), AddedSafetyChecks(false) {}
269
270 // Perform the actual loop widening (vectorization).
271 void vectorize(LoopVectorizationLegality *L) {
272 Legal = L;
273 // Create a new empty loop. Unlink the old loop and connect the new one.
274 createEmptyLoop();
275 // Widen each instruction in the old loop to a new one in the new loop.
276 // Use the Legality module to find the induction and reduction variables.
277 vectorizeLoop();
278 // Register the new loop and update the analysis passes.
279 updateAnalysis();
280 }
281
282 // Return true if any runtime check is added.
283 bool IsSafetyChecksAdded() {
284 return AddedSafetyChecks;
285 }
286
287 virtual ~InnerLoopVectorizer() {}
288
289protected:
290 /// A small list of PHINodes.
291 typedef SmallVector<PHINode*, 4> PhiVector;
292 /// When we unroll loops we have multiple vector values for each scalar.
293 /// This data structure holds the unrolled and vectorized values that
294 /// originated from one scalar instruction.
295 typedef SmallVector<Value*, 2> VectorParts;
296
297 // When we if-convert we need to create edge masks. We have to cache values
298 // so that we don't end up with exponential recursion/IR.
299 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
300 VectorParts> EdgeMaskCache;
301
302 /// \brief Add checks for strides that were assumed to be 1.
303 ///
304 /// Returns the last check instruction and the first check instruction in the
305 /// pair as (first, last).
306 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
307
308 /// Create an empty loop, based on the loop ranges of the old loop.
309 void createEmptyLoop();
310 /// Copy and widen the instructions from the old loop.
311 virtual void vectorizeLoop();
312
313 /// \brief The Loop exit block may have single value PHI nodes where the
314 /// incoming value is 'Undef'. While vectorizing we only handled real values
315 /// that were defined inside the loop. Here we fix the 'undef case'.
316 /// See PR14725.
317 void fixLCSSAPHIs();
318
319 /// A helper function that computes the predicate of the block BB, assuming
320 /// that the header block of the loop is set to True. It returns the *entry*
321 /// mask for the block BB.
322 VectorParts createBlockInMask(BasicBlock *BB);
323 /// A helper function that computes the predicate of the edge between SRC
324 /// and DST.
325 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
326
327 /// A helper function to vectorize a single BB within the innermost loop.
328 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
329
330 /// Vectorize a single PHINode in a block. This method handles the induction
331 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
332 /// arbitrary length vectors.
333 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
334 unsigned UF, unsigned VF, PhiVector *PV);
335
336 /// Insert the new loop to the loop hierarchy and pass manager
337 /// and update the analysis passes.
338 void updateAnalysis();
339
340 /// This instruction is un-vectorizable. Implement it as a sequence
341 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
342 /// scalarized instruction behind an if block predicated on the control
343 /// dependence of the instruction.
344 virtual void scalarizeInstruction(Instruction *Instr,
345 bool IfPredicateStore=false);
346
347 /// Vectorize Load and Store instructions,
348 virtual void vectorizeMemoryInstruction(Instruction *Instr);
349
350 /// Create a broadcast instruction. This method generates a broadcast
351 /// instruction (shuffle) for loop invariant values and for the induction
352 /// value. If this is the induction variable then we extend it to N, N+1, ...
353 /// this is needed because each iteration in the loop corresponds to a SIMD
354 /// element.
355 virtual Value *getBroadcastInstrs(Value *V);
356
357 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
358 /// to each vector element of Val. The sequence starts at StartIndex.
359 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
360
361 /// When we go over instructions in the basic block we rely on previous
362 /// values within the current basic block or on loop invariant values.
363 /// When we widen (vectorize) values we place them in the map. If the values
364 /// are not within the map, they have to be loop invariant, so we simply
365 /// broadcast them into a vector.
366 VectorParts &getVectorValue(Value *V);
367
368 /// Try to vectorize the interleaved access group that \p Instr belongs to.
369 void vectorizeInterleaveGroup(Instruction *Instr);
370
371 /// Generate a shuffle sequence that will reverse the vector Vec.
372 virtual Value *reverseVector(Value *Vec);
373
374 /// This is a helper class that holds the vectorizer state. It maps scalar
375 /// instructions to vector instructions. When the code is 'unrolled' then
376 /// then a single scalar value is mapped to multiple vector parts. The parts
377 /// are stored in the VectorPart type.
378 struct ValueMap {
379 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
380 /// are mapped.
381 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
382
383 /// \return True if 'Key' is saved in the Value Map.
384 bool has(Value *Key) const { return MapStorage.count(Key); }
385
386 /// Initializes a new entry in the map. Sets all of the vector parts to the
387 /// save value in 'Val'.
388 /// \return A reference to a vector with splat values.
389 VectorParts &splat(Value *Key, Value *Val) {
390 VectorParts &Entry = MapStorage[Key];
391 Entry.assign(UF, Val);
392 return Entry;
393 }
394
395 ///\return A reference to the value that is stored at 'Key'.
396 VectorParts &get(Value *Key) {
397 VectorParts &Entry = MapStorage[Key];
398 if (Entry.empty())
399 Entry.resize(UF);
400 assert(Entry.size() == UF)((Entry.size() == UF) ? static_cast<void> (0) : __assert_fail
("Entry.size() == UF", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 400, __PRETTY_FUNCTION__));
401 return Entry;
402 }
403
404 private:
405 /// The unroll factor. Each entry in the map stores this number of vector
406 /// elements.
407 unsigned UF;
408
409 /// Map storage. We use std::map and not DenseMap because insertions to a
410 /// dense map invalidates its iterators.
411 std::map<Value *, VectorParts> MapStorage;
412 };
413
414 /// The original loop.
415 Loop *OrigLoop;
416 /// Scev analysis to use.
417 ScalarEvolution *SE;
418 /// Loop Info.
419 LoopInfo *LI;
420 /// Dominator Tree.
421 DominatorTree *DT;
422 /// Alias Analysis.
423 AliasAnalysis *AA;
424 /// Target Library Info.
425 const TargetLibraryInfo *TLI;
426 /// Target Transform Info.
427 const TargetTransformInfo *TTI;
428
429 /// The vectorization SIMD factor to use. Each vector will have this many
430 /// vector elements.
431 unsigned VF;
432
433protected:
434 /// The vectorization unroll factor to use. Each scalar is vectorized to this
435 /// many different vector instructions.
436 unsigned UF;
437
438 /// The builder that we use
439 IRBuilder<> Builder;
440
441 // --- Vectorization state ---
442
443 /// The vector-loop preheader.
444 BasicBlock *LoopVectorPreHeader;
445 /// The scalar-loop preheader.
446 BasicBlock *LoopScalarPreHeader;
447 /// Middle Block between the vector and the scalar.
448 BasicBlock *LoopMiddleBlock;
449 ///The ExitBlock of the scalar loop.
450 BasicBlock *LoopExitBlock;
451 ///The vector loop body.
452 SmallVector<BasicBlock *, 4> LoopVectorBody;
453 ///The scalar loop body.
454 BasicBlock *LoopScalarBody;
455 /// A list of all bypass blocks. The first block is the entry of the loop.
456 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
457
458 /// The new Induction variable which was added to the new block.
459 PHINode *Induction;
460 /// The induction variable of the old basic block.
461 PHINode *OldInduction;
462 /// Holds the extended (to the widest induction type) start index.
463 Value *ExtendedIdx;
464 /// Maps scalars to widened vectors.
465 ValueMap WidenMap;
466 EdgeMaskCache MaskCache;
467
468 LoopVectorizationLegality *Legal;
469
470 // Record whether runtime check is added.
471 bool AddedSafetyChecks;
472};
473
474class InnerLoopUnroller : public InnerLoopVectorizer {
475public:
476 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
477 DominatorTree *DT, const TargetLibraryInfo *TLI,
478 const TargetTransformInfo *TTI, unsigned UnrollFactor)
479 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
480
481private:
482 void scalarizeInstruction(Instruction *Instr,
483 bool IfPredicateStore = false) override;
484 void vectorizeMemoryInstruction(Instruction *Instr) override;
485 Value *getBroadcastInstrs(Value *V) override;
486 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
487 Value *reverseVector(Value *Vec) override;
488};
489
490/// \brief Look for a meaningful debug location on the instruction or it's
491/// operands.
492static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
493 if (!I)
494 return I;
495
496 DebugLoc Empty;
497 if (I->getDebugLoc() != Empty)
498 return I;
499
500 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
501 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
502 if (OpInst->getDebugLoc() != Empty)
503 return OpInst;
504 }
505
506 return I;
507}
508
509/// \brief Set the debug location in the builder using the debug location in the
510/// instruction.
511static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
512 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
513 B.SetCurrentDebugLocation(Inst->getDebugLoc());
514 else
515 B.SetCurrentDebugLocation(DebugLoc());
516}
517
518#ifndef NDEBUG
519/// \return string containing a file name and a line # for the given loop.
520static std::string getDebugLocString(const Loop *L) {
521 std::string Result;
522 if (L) {
523 raw_string_ostream OS(Result);
524 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
525 LoopDbgLoc.print(OS);
526 else
527 // Just print the module name.
528 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
529 OS.flush();
530 }
531 return Result;
532}
533#endif
534
535/// \brief Propagate known metadata from one instruction to another.
536static void propagateMetadata(Instruction *To, const Instruction *From) {
537 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
538 From->getAllMetadataOtherThanDebugLoc(Metadata);
539
540 for (auto M : Metadata) {
541 unsigned Kind = M.first;
542
543 // These are safe to transfer (this is safe for TBAA, even when we
544 // if-convert, because should that metadata have had a control dependency
545 // on the condition, and thus actually aliased with some other
546 // non-speculated memory access when the condition was false, this would be
547 // caught by the runtime overlap checks).
548 if (Kind != LLVMContext::MD_tbaa &&
549 Kind != LLVMContext::MD_alias_scope &&
550 Kind != LLVMContext::MD_noalias &&
551 Kind != LLVMContext::MD_fpmath)
552 continue;
553
554 To->setMetadata(Kind, M.second);
555 }
556}
557
558/// \brief Propagate known metadata from one instruction to a vector of others.
559static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
560 for (Value *V : To)
561 if (Instruction *I = dyn_cast<Instruction>(V))
562 propagateMetadata(I, From);
563}
564
565/// \brief The group of interleaved loads/stores sharing the same stride and
566/// close to each other.
567///
568/// Each member in this group has an index starting from 0, and the largest
569/// index should be less than interleaved factor, which is equal to the absolute
570/// value of the access's stride.
571///
572/// E.g. An interleaved load group of factor 4:
573/// for (unsigned i = 0; i < 1024; i+=4) {
574/// a = A[i]; // Member of index 0
575/// b = A[i+1]; // Member of index 1
576/// d = A[i+3]; // Member of index 3
577/// ...
578/// }
579///
580/// An interleaved store group of factor 4:
581/// for (unsigned i = 0; i < 1024; i+=4) {
582/// ...
583/// A[i] = a; // Member of index 0
584/// A[i+1] = b; // Member of index 1
585/// A[i+2] = c; // Member of index 2
586/// A[i+3] = d; // Member of index 3
587/// }
588///
589/// Note: the interleaved load group could have gaps (missing members), but
590/// the interleaved store group doesn't allow gaps.
591class InterleaveGroup {
592public:
593 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
594 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
595 assert(Align && "The alignment should be non-zero")((Align && "The alignment should be non-zero") ? static_cast
<void> (0) : __assert_fail ("Align && \"The alignment should be non-zero\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 595, __PRETTY_FUNCTION__));
596
597 Factor = std::abs(Stride);
598 assert(Factor > 1 && "Invalid interleave factor")((Factor > 1 && "Invalid interleave factor") ? static_cast
<void> (0) : __assert_fail ("Factor > 1 && \"Invalid interleave factor\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 598, __PRETTY_FUNCTION__));
599
600 Reverse = Stride < 0;
601 Members[0] = Instr;
602 }
603
604 bool isReverse() const { return Reverse; }
605 unsigned getFactor() const { return Factor; }
606 unsigned getAlignment() const { return Align; }
607 unsigned getNumMembers() const { return Members.size(); }
608
609 /// \brief Try to insert a new member \p Instr with index \p Index and
610 /// alignment \p NewAlign. The index is related to the leader and it could be
611 /// negative if it is the new leader.
612 ///
613 /// \returns false if the instruction doesn't belong to the group.
614 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
615 assert(NewAlign && "The new member's alignment should be non-zero")((NewAlign && "The new member's alignment should be non-zero"
) ? static_cast<void> (0) : __assert_fail ("NewAlign && \"The new member's alignment should be non-zero\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 615, __PRETTY_FUNCTION__));
616
617 int Key = Index + SmallestKey;
618
619 // Skip if there is already a member with the same index.
620 if (Members.count(Key))
621 return false;
622
623 if (Key > LargestKey) {
624 // The largest index is always less than the interleave factor.
625 if (Index >= static_cast<int>(Factor))
626 return false;
627
628 LargestKey = Key;
629 } else if (Key < SmallestKey) {
630 // The largest index is always less than the interleave factor.
631 if (LargestKey - Key >= static_cast<int>(Factor))
632 return false;
633
634 SmallestKey = Key;
635 }
636
637 // It's always safe to select the minimum alignment.
638 Align = std::min(Align, NewAlign);
639 Members[Key] = Instr;
640 return true;
641 }
642
643 /// \brief Get the member with the given index \p Index
644 ///
645 /// \returns nullptr if contains no such member.
646 Instruction *getMember(unsigned Index) const {
647 int Key = SmallestKey + Index;
648 if (!Members.count(Key))
649 return nullptr;
650
651 return Members.find(Key)->second;
652 }
653
654 /// \brief Get the index for the given member. Unlike the key in the member
655 /// map, the index starts from 0.
656 unsigned getIndex(Instruction *Instr) const {
657 for (auto I : Members)
658 if (I.second == Instr)
659 return I.first - SmallestKey;
660
661 llvm_unreachable("InterleaveGroup contains no such member")::llvm::llvm_unreachable_internal("InterleaveGroup contains no such member"
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 661);
662 }
663
664 Instruction *getInsertPos() const { return InsertPos; }
665 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
666
667private:
668 unsigned Factor; // Interleave Factor.
669 bool Reverse;
670 unsigned Align;
671 DenseMap<int, Instruction *> Members;
672 int SmallestKey;
673 int LargestKey;
674
675 // To avoid breaking dependences, vectorized instructions of an interleave
676 // group should be inserted at either the first load or the last store in
677 // program order.
678 //
679 // E.g. %even = load i32 // Insert Position
680 // %add = add i32 %even // Use of %even
681 // %odd = load i32
682 //
683 // store i32 %even
684 // %odd = add i32 // Def of %odd
685 // store i32 %odd // Insert Position
686 Instruction *InsertPos;
687};
688
689/// \brief Drive the analysis of interleaved memory accesses in the loop.
690///
691/// Use this class to analyze interleaved accesses only when we can vectorize
692/// a loop. Otherwise it's meaningless to do analysis as the vectorization
693/// on interleaved accesses is unsafe.
694///
695/// The analysis collects interleave groups and records the relationships
696/// between the member and the group in a map.
697class InterleavedAccessInfo {
698public:
699 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
700 : SE(SE), TheLoop(L), DT(DT) {}
701
702 ~InterleavedAccessInfo() {
703 SmallSet<InterleaveGroup *, 4> DelSet;
704 // Avoid releasing a pointer twice.
705 for (auto &I : InterleaveGroupMap)
706 DelSet.insert(I.second);
707 for (auto *Ptr : DelSet)
708 delete Ptr;
709 }
710
711 /// \brief Analyze the interleaved accesses and collect them in interleave
712 /// groups. Substitute symbolic strides using \p Strides.
713 void analyzeInterleaving(const ValueToValueMap &Strides);
714
715 /// \brief Check if \p Instr belongs to any interleave group.
716 bool isInterleaved(Instruction *Instr) const {
717 return InterleaveGroupMap.count(Instr);
718 }
719
720 /// \brief Get the interleave group that \p Instr belongs to.
721 ///
722 /// \returns nullptr if doesn't have such group.
723 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
724 if (InterleaveGroupMap.count(Instr))
725 return InterleaveGroupMap.find(Instr)->second;
726 return nullptr;
727 }
728
729private:
730 ScalarEvolution *SE;
731 Loop *TheLoop;
732 DominatorTree *DT;
733
734 /// Holds the relationships between the members and the interleave group.
735 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
736
737 /// \brief The descriptor for a strided memory access.
738 struct StrideDescriptor {
739 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
740 unsigned Align)
741 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
742
743 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
744
745 int Stride; // The access's stride. It is negative for a reverse access.
746 const SCEV *Scev; // The scalar expression of this access
747 unsigned Size; // The size of the memory object.
748 unsigned Align; // The alignment of this access.
749 };
750
751 /// \brief Create a new interleave group with the given instruction \p Instr,
752 /// stride \p Stride and alignment \p Align.
753 ///
754 /// \returns the newly created interleave group.
755 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
756 unsigned Align) {
757 assert(!InterleaveGroupMap.count(Instr) &&((!InterleaveGroupMap.count(Instr) && "Already in an interleaved access group"
) ? static_cast<void> (0) : __assert_fail ("!InterleaveGroupMap.count(Instr) && \"Already in an interleaved access group\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 758, __PRETTY_FUNCTION__))
758 "Already in an interleaved access group")((!InterleaveGroupMap.count(Instr) && "Already in an interleaved access group"
) ? static_cast<void> (0) : __assert_fail ("!InterleaveGroupMap.count(Instr) && \"Already in an interleaved access group\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 758, __PRETTY_FUNCTION__));
759 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
760 return InterleaveGroupMap[Instr];
761 }
762
763 /// \brief Release the group and remove all the relationships.
764 void releaseGroup(InterleaveGroup *Group) {
765 for (unsigned i = 0; i < Group->getFactor(); i++)
766 if (Instruction *Member = Group->getMember(i))
767 InterleaveGroupMap.erase(Member);
768
769 delete Group;
770 }
771
772 /// \brief Collect all the accesses with a constant stride in program order.
773 void collectConstStridedAccesses(
774 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
775 const ValueToValueMap &Strides);
776};
777
778/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
779/// to what vectorization factor.
780/// This class does not look at the profitability of vectorization, only the
781/// legality. This class has two main kinds of checks:
782/// * Memory checks - The code in canVectorizeMemory checks if vectorization
783/// will change the order of memory accesses in a way that will change the
784/// correctness of the program.
785/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
786/// checks for a number of different conditions, such as the availability of a
787/// single induction variable, that all types are supported and vectorize-able,
788/// etc. This code reflects the capabilities of InnerLoopVectorizer.
789/// This class is also used by InnerLoopVectorizer for identifying
790/// induction variable and the different reduction variables.
791class LoopVectorizationLegality {
792public:
793 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
794 TargetLibraryInfo *TLI, AliasAnalysis *AA,
795 Function *F, const TargetTransformInfo *TTI,
796 LoopAccessAnalysis *LAA)
797 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
798 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
799 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
800
801 /// This enum represents the kinds of inductions that we support.
802 enum InductionKind {
803 IK_NoInduction, ///< Not an induction variable.
804 IK_IntInduction, ///< Integer induction variable. Step = C.
805 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
806 };
807
808 /// A struct for saving information about induction variables.
809 struct InductionInfo {
810 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
811 : StartValue(Start), IK(K), StepValue(Step) {
812 assert(IK != IK_NoInduction && "Not an induction")((IK != IK_NoInduction && "Not an induction") ? static_cast
<void> (0) : __assert_fail ("IK != IK_NoInduction && \"Not an induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 812, __PRETTY_FUNCTION__));
813 assert(StartValue && "StartValue is null")((StartValue && "StartValue is null") ? static_cast<
void> (0) : __assert_fail ("StartValue && \"StartValue is null\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 813, __PRETTY_FUNCTION__));
814 assert(StepValue && !StepValue->isZero() && "StepValue is zero")((StepValue && !StepValue->isZero() && "StepValue is zero"
) ? static_cast<void> (0) : __assert_fail ("StepValue && !StepValue->isZero() && \"StepValue is zero\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 814, __PRETTY_FUNCTION__));
815 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&(((IK != IK_PtrInduction || StartValue->getType()->isPointerTy
()) && "StartValue is not a pointer for pointer induction"
) ? static_cast<void> (0) : __assert_fail ("(IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) && \"StartValue is not a pointer for pointer induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 816, __PRETTY_FUNCTION__))
816 "StartValue is not a pointer for pointer induction")(((IK != IK_PtrInduction || StartValue->getType()->isPointerTy
()) && "StartValue is not a pointer for pointer induction"
) ? static_cast<void> (0) : __assert_fail ("(IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) && \"StartValue is not a pointer for pointer induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 816, __PRETTY_FUNCTION__));
817 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&(((IK != IK_IntInduction || StartValue->getType()->isIntegerTy
()) && "StartValue is not an integer for integer induction"
) ? static_cast<void> (0) : __assert_fail ("(IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) && \"StartValue is not an integer for integer induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 818, __PRETTY_FUNCTION__))
818 "StartValue is not an integer for integer induction")(((IK != IK_IntInduction || StartValue->getType()->isIntegerTy
()) && "StartValue is not an integer for integer induction"
) ? static_cast<void> (0) : __assert_fail ("(IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) && \"StartValue is not an integer for integer induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 818, __PRETTY_FUNCTION__));
819 assert(StepValue->getType()->isIntegerTy() &&((StepValue->getType()->isIntegerTy() && "StepValue is not an integer"
) ? static_cast<void> (0) : __assert_fail ("StepValue->getType()->isIntegerTy() && \"StepValue is not an integer\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 820, __PRETTY_FUNCTION__))
820 "StepValue is not an integer")((StepValue->getType()->isIntegerTy() && "StepValue is not an integer"
) ? static_cast<void> (0) : __assert_fail ("StepValue->getType()->isIntegerTy() && \"StepValue is not an integer\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 820, __PRETTY_FUNCTION__));
821 }
822 InductionInfo()
823 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
824
825 /// Get the consecutive direction. Returns:
826 /// 0 - unknown or non-consecutive.
827 /// 1 - consecutive and increasing.
828 /// -1 - consecutive and decreasing.
829 int getConsecutiveDirection() const {
830 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
831 return StepValue->getSExtValue();
832 return 0;
833 }
834
835 /// Compute the transformed value of Index at offset StartValue using step
836 /// StepValue.
837 /// For integer induction, returns StartValue + Index * StepValue.
838 /// For pointer induction, returns StartValue[Index * StepValue].
839 /// FIXME: The newly created binary instructions should contain nsw/nuw
840 /// flags, which can be found from the original scalar operations.
841 Value *transform(IRBuilder<> &B, Value *Index) const {
842 switch (IK) {
843 case IK_IntInduction:
844 assert(Index->getType() == StartValue->getType() &&((Index->getType() == StartValue->getType() && "Index type does not match StartValue type"
) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 845, __PRETTY_FUNCTION__))
845 "Index type does not match StartValue type")((Index->getType() == StartValue->getType() && "Index type does not match StartValue type"
) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 845, __PRETTY_FUNCTION__));
846 if (StepValue->isMinusOne())
847 return B.CreateSub(StartValue, Index);
848 if (!StepValue->isOne())
849 Index = B.CreateMul(Index, StepValue);
850 return B.CreateAdd(StartValue, Index);
851
852 case IK_PtrInduction:
853 assert(Index->getType() == StepValue->getType() &&((Index->getType() == StepValue->getType() && "Index type does not match StepValue type"
) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StepValue->getType() && \"Index type does not match StepValue type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 854, __PRETTY_FUNCTION__))
854 "Index type does not match StepValue type")((Index->getType() == StepValue->getType() && "Index type does not match StepValue type"
) ? static_cast<void> (0) : __assert_fail ("Index->getType() == StepValue->getType() && \"Index type does not match StepValue type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 854, __PRETTY_FUNCTION__));
855 if (StepValue->isMinusOne())
856 Index = B.CreateNeg(Index);
857 else if (!StepValue->isOne())
858 Index = B.CreateMul(Index, StepValue);
859 return B.CreateGEP(nullptr, StartValue, Index);
860
861 case IK_NoInduction:
862 return nullptr;
863 }
864 llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 864);
865 }
866
867 /// Start value.
868 TrackingVH<Value> StartValue;
869 /// Induction kind.
870 InductionKind IK;
871 /// Step value.
872 ConstantInt *StepValue;
873 };
874
875 /// ReductionList contains the reduction descriptors for all
876 /// of the reductions that were found in the loop.
877 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
878
879 /// InductionList saves induction variables and maps them to the
880 /// induction descriptor.
881 typedef MapVector<PHINode*, InductionInfo> InductionList;
882
883 /// Returns true if it is legal to vectorize this loop.
884 /// This does not mean that it is profitable to vectorize this
885 /// loop, only that it is legal to do so.
886 bool canVectorize();
887
888 /// Returns the Induction variable.
889 PHINode *getInduction() { return Induction; }
890
891 /// Returns the reduction variables found in the loop.
892 ReductionList *getReductionVars() { return &Reductions; }
893
894 /// Returns the induction variables found in the loop.
895 InductionList *getInductionVars() { return &Inductions; }
896
897 /// Returns the widest induction type.
898 Type *getWidestInductionType() { return WidestIndTy; }
899
900 /// Returns True if V is an induction variable in this loop.
901 bool isInductionVariable(const Value *V);
902
903 /// Return true if the block BB needs to be predicated in order for the loop
904 /// to be vectorized.
905 bool blockNeedsPredication(BasicBlock *BB);
906
907 /// Check if this pointer is consecutive when vectorizing. This happens
908 /// when the last index of the GEP is the induction variable, or that the
909 /// pointer itself is an induction variable.
910 /// This check allows us to vectorize A[idx] into a wide load/store.
911 /// Returns:
912 /// 0 - Stride is unknown or non-consecutive.
913 /// 1 - Address is consecutive.
914 /// -1 - Address is consecutive, and decreasing.
915 int isConsecutivePtr(Value *Ptr);
916
917 /// Returns true if the value V is uniform within the loop.
918 bool isUniform(Value *V);
919
920 /// Returns true if this instruction will remain scalar after vectorization.
921 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
922
923 /// Returns the information that we collected about runtime memory check.
924 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
925 return LAI->getRuntimePointerCheck();
926 }
927
928 const LoopAccessInfo *getLAI() const {
929 return LAI;
930 }
931
932 /// \brief Check if \p Instr belongs to any interleaved access group.
933 bool isAccessInterleaved(Instruction *Instr) {
934 return InterleaveInfo.isInterleaved(Instr);
935 }
936
937 /// \brief Get the interleaved access group that \p Instr belongs to.
938 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
939 return InterleaveInfo.getInterleaveGroup(Instr);
940 }
941
942 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
943
944 bool hasStride(Value *V) { return StrideSet.count(V); }
945 bool mustCheckStrides() { return !StrideSet.empty(); }
946 SmallPtrSet<Value *, 8>::iterator strides_begin() {
947 return StrideSet.begin();
948 }
949 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
950
951 /// Returns true if the target machine supports masked store operation
952 /// for the given \p DataType and kind of access to \p Ptr.
953 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
954 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
955 }
956 /// Returns true if the target machine supports masked load operation
957 /// for the given \p DataType and kind of access to \p Ptr.
958 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
959 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
960 }
961 /// Returns true if vector representation of the instruction \p I
962 /// requires mask.
963 bool isMaskRequired(const Instruction* I) {
964 return (MaskedOp.count(I) != 0);
965 }
966 unsigned getNumStores() const {
967 return LAI->getNumStores();
968 }
969 unsigned getNumLoads() const {
970 return LAI->getNumLoads();
971 }
972 unsigned getNumPredStores() const {
973 return NumPredStores;
974 }
975private:
976 /// Check if a single basic block loop is vectorizable.
977 /// At this point we know that this is a loop with a constant trip count
978 /// and we only need to check individual instructions.
979 bool canVectorizeInstrs();
980
981 /// When we vectorize loops we may change the order in which
982 /// we read and write from memory. This method checks if it is
983 /// legal to vectorize the code, considering only memory constrains.
984 /// Returns true if the loop is vectorizable
985 bool canVectorizeMemory();
986
987 /// Return true if we can vectorize this loop using the IF-conversion
988 /// transformation.
989 bool canVectorizeWithIfConvert();
990
991 /// Collect the variables that need to stay uniform after vectorization.
992 void collectLoopUniforms();
993
994 /// Return true if all of the instructions in the block can be speculatively
995 /// executed. \p SafePtrs is a list of addresses that are known to be legal
996 /// and we know that we can read from them without segfault.
997 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
998
999 /// Returns the induction kind of Phi and record the step. This function may
1000 /// return NoInduction if the PHI is not an induction variable.
1001 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
1002
1003 /// \brief Collect memory access with loop invariant strides.
1004 ///
1005 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1006 /// invariant.
1007 void collectStridedAccess(Value *LoadOrStoreInst);
1008
1009 /// Report an analysis message to assist the user in diagnosing loops that are
1010 /// not vectorized. These are handled as LoopAccessReport rather than
1011 /// VectorizationReport because the << operator of VectorizationReport returns
1012 /// LoopAccessReport.
1013 void emitAnalysis(const LoopAccessReport &Message) {
1014 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME"loop-vectorize");
1015 }
1016
1017 unsigned NumPredStores;
1018
1019 /// The loop that we evaluate.
1020 Loop *TheLoop;
1021 /// Scev analysis.
1022 ScalarEvolution *SE;
1023 /// Target Library Info.
1024 TargetLibraryInfo *TLI;
1025 /// Parent function
1026 Function *TheFunction;
1027 /// Target Transform Info
1028 const TargetTransformInfo *TTI;
1029 /// Dominator Tree.
1030 DominatorTree *DT;
1031 // LoopAccess analysis.
1032 LoopAccessAnalysis *LAA;
1033 // And the loop-accesses info corresponding to this loop. This pointer is
1034 // null until canVectorizeMemory sets it up.
1035 const LoopAccessInfo *LAI;
1036
1037 /// The interleave access information contains groups of interleaved accesses
1038 /// with the same stride and close to each other.
1039 InterleavedAccessInfo InterleaveInfo;
1040
1041 // --- vectorization state --- //
1042
1043 /// Holds the integer induction variable. This is the counter of the
1044 /// loop.
1045 PHINode *Induction;
1046 /// Holds the reduction variables.
1047 ReductionList Reductions;
1048 /// Holds all of the induction variables that we found in the loop.
1049 /// Notice that inductions don't need to start at zero and that induction
1050 /// variables can be pointers.
1051 InductionList Inductions;
1052 /// Holds the widest induction type encountered.
1053 Type *WidestIndTy;
1054
1055 /// Allowed outside users. This holds the reduction
1056 /// vars which can be accessed from outside the loop.
1057 SmallPtrSet<Value*, 4> AllowedExit;
1058 /// This set holds the variables which are known to be uniform after
1059 /// vectorization.
1060 SmallPtrSet<Instruction*, 4> Uniforms;
1061
1062 /// Can we assume the absence of NaNs.
1063 bool HasFunNoNaNAttr;
1064
1065 ValueToValueMap Strides;
1066 SmallPtrSet<Value *, 8> StrideSet;
1067
1068 /// While vectorizing these instructions we have to generate a
1069 /// call to the appropriate masked intrinsic
1070 SmallPtrSet<const Instruction*, 8> MaskedOp;
1071};
1072
1073/// LoopVectorizationCostModel - estimates the expected speedups due to
1074/// vectorization.
1075/// In many cases vectorization is not profitable. This can happen because of
1076/// a number of reasons. In this class we mainly attempt to predict the
1077/// expected speedup/slowdowns due to the supported instruction set. We use the
1078/// TargetTransformInfo to query the different backends for the cost of
1079/// different operations.
1080class LoopVectorizationCostModel {
1081public:
1082 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1083 LoopVectorizationLegality *Legal,
1084 const TargetTransformInfo &TTI,
1085 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1086 const Function *F, const LoopVectorizeHints *Hints)
1087 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1088 TheFunction(F), Hints(Hints) {
1089 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
1090 }
1091
1092 /// Information about vectorization costs
1093 struct VectorizationFactor {
1094 unsigned Width; // Vector width with best cost
1095 unsigned Cost; // Cost of the loop with that width
1096 };
1097 /// \return The most profitable vectorization factor and the cost of that VF.
1098 /// This method checks every power of two up to VF. If UserVF is not ZERO
1099 /// then this vectorization factor will be selected if vectorization is
1100 /// possible.
1101 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1102
1103 /// \return The size (in bits) of the widest type in the code that
1104 /// needs to be vectorized. We ignore values that remain scalar such as
1105 /// 64 bit loop indices.
1106 unsigned getWidestType();
1107
1108 /// \return The most profitable unroll factor.
1109 /// If UserUF is non-zero then this method finds the best unroll-factor
1110 /// based on register pressure and other parameters.
1111 /// VF and LoopCost are the selected vectorization factor and the cost of the
1112 /// selected VF.
1113 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
1114
1115 /// \brief A struct that represents some properties of the register usage
1116 /// of a loop.
1117 struct RegisterUsage {
1118 /// Holds the number of loop invariant values that are used in the loop.
1119 unsigned LoopInvariantRegs;
1120 /// Holds the maximum number of concurrent live intervals in the loop.
1121 unsigned MaxLocalUsers;
1122 /// Holds the number of instructions in the loop.
1123 unsigned NumInstructions;
1124 };
1125
1126 /// \return information about the register usage of the loop.
1127 RegisterUsage calculateRegisterUsage();
1128
1129private:
1130 /// Returns the expected execution cost. The unit of the cost does
1131 /// not matter because we use the 'cost' units to compare different
1132 /// vector widths. The cost that is returned is *not* normalized by
1133 /// the factor width.
1134 unsigned expectedCost(unsigned VF);
1135
1136 /// Returns the execution time cost of an instruction for a given vector
1137 /// width. Vector width of one means scalar.
1138 unsigned getInstructionCost(Instruction *I, unsigned VF);
1139
1140 /// Returns whether the instruction is a load or store and will be a emitted
1141 /// as a vector operation.
1142 bool isConsecutiveLoadOrStore(Instruction *I);
1143
1144 /// Report an analysis message to assist the user in diagnosing loops that are
1145 /// not vectorized. These are handled as LoopAccessReport rather than
1146 /// VectorizationReport because the << operator of VectorizationReport returns
1147 /// LoopAccessReport.
1148 void emitAnalysis(const LoopAccessReport &Message) {
1149 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME"loop-vectorize");
1150 }
1151
1152 /// Values used only by @llvm.assume calls.
1153 SmallPtrSet<const Value *, 32> EphValues;
1154
1155 /// The loop that we evaluate.
1156 Loop *TheLoop;
1157 /// Scev analysis.
1158 ScalarEvolution *SE;
1159 /// Loop Info analysis.
1160 LoopInfo *LI;
1161 /// Vectorization legality.
1162 LoopVectorizationLegality *Legal;
1163 /// Vector target information.
1164 const TargetTransformInfo &TTI;
1165 /// Target Library Info.
1166 const TargetLibraryInfo *TLI;
1167 const Function *TheFunction;
1168 // Loop Vectorize Hint.
1169 const LoopVectorizeHints *Hints;
1170};
1171
1172/// Utility class for getting and setting loop vectorizer hints in the form
1173/// of loop metadata.
1174/// This class keeps a number of loop annotations locally (as member variables)
1175/// and can, upon request, write them back as metadata on the loop. It will
1176/// initially scan the loop for existing metadata, and will update the local
1177/// values based on information in the loop.
1178/// We cannot write all values to metadata, as the mere presence of some info,
1179/// for example 'force', means a decision has been made. So, we need to be
1180/// careful NOT to add them if the user hasn't specifically asked so.
1181class LoopVectorizeHints {
1182 enum HintKind {
1183 HK_WIDTH,
1184 HK_UNROLL,
1185 HK_FORCE
1186 };
1187
1188 /// Hint - associates name and validation with the hint value.
1189 struct Hint {
1190 const char * Name;
1191 unsigned Value; // This may have to change for non-numeric values.
1192 HintKind Kind;
1193
1194 Hint(const char * Name, unsigned Value, HintKind Kind)
1195 : Name(Name), Value(Value), Kind(Kind) { }
1196
1197 bool validate(unsigned Val) {
1198 switch (Kind) {
1199 case HK_WIDTH:
1200 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1201 case HK_UNROLL:
1202 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1203 case HK_FORCE:
1204 return (Val <= 1);
1205 }
1206 return false;
1207 }
1208 };
1209
1210 /// Vectorization width.
1211 Hint Width;
1212 /// Vectorization interleave factor.
1213 Hint Interleave;
1214 /// Vectorization forced
1215 Hint Force;
1216
1217 /// Return the loop metadata prefix.
1218 static StringRef Prefix() { return "llvm.loop."; }
1219
1220public:
1221 enum ForceKind {
1222 FK_Undefined = -1, ///< Not selected.
1223 FK_Disabled = 0, ///< Forcing disabled.
1224 FK_Enabled = 1, ///< Forcing enabled.
1225 };
1226
1227 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1228 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1229 HK_WIDTH),
1230 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1231 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1232 TheLoop(L) {
1233 // Populate values with existing loop metadata.
1234 getHintsFromMetadata();
1235
1236 // force-vector-interleave overrides DisableInterleaving.
1237 if (VectorizerParams::isInterleaveForced())
1238 Interleave.Value = VectorizerParams::VectorizationInterleave;
1239
1240 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { if (DisableInterleaving && Interleave
.Value == 1) dbgs() << "LV: Interleaving disabled by the pass manager\n"
; } } while (0)
1241 << "LV: Interleaving disabled by the pass manager\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { if (DisableInterleaving && Interleave
.Value == 1) dbgs() << "LV: Interleaving disabled by the pass manager\n"
; } } while (0);
1242 }
1243
1244 /// Mark the loop L as already vectorized by setting the width to 1.
1245 void setAlreadyVectorized() {
1246 Width.Value = Interleave.Value = 1;
1247 Hint Hints[] = {Width, Interleave};
1248 writeHintsToMetadata(Hints);
1249 }
1250
1251 /// Dumps all the hint information.
1252 std::string emitRemark() const {
1253 VectorizationReport R;
1254 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1255 R << "vectorization is explicitly disabled";
1256 else {
1257 R << "use -Rpass-analysis=loop-vectorize for more info";
1258 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1259 R << " (Force=true";
1260 if (Width.Value != 0)
1261 R << ", Vector Width=" << Width.Value;
1262 if (Interleave.Value != 0)
1263 R << ", Interleave Count=" << Interleave.Value;
1264 R << ")";
1265 }
1266 }
1267
1268 return R.str();
1269 }
1270
1271 unsigned getWidth() const { return Width.Value; }
1272 unsigned getInterleave() const { return Interleave.Value; }
1273 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1274
1275private:
1276 /// Find hints specified in the loop metadata and update local values.
1277 void getHintsFromMetadata() {
1278 MDNode *LoopID = TheLoop->getLoopID();
1279 if (!LoopID)
1280 return;
1281
1282 // First operand should refer to the loop id itself.
1283 assert(LoopID->getNumOperands() > 0 && "requires at least one operand")((LoopID->getNumOperands() > 0 && "requires at least one operand"
) ? static_cast<void> (0) : __assert_fail ("LoopID->getNumOperands() > 0 && \"requires at least one operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1283, __PRETTY_FUNCTION__));
1284 assert(LoopID->getOperand(0) == LoopID && "invalid loop id")((LoopID->getOperand(0) == LoopID && "invalid loop id"
) ? static_cast<void> (0) : __assert_fail ("LoopID->getOperand(0) == LoopID && \"invalid loop id\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1284, __PRETTY_FUNCTION__));
1285
1286 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1287 const MDString *S = nullptr;
1288 SmallVector<Metadata *, 4> Args;
1289
1290 // The expected hint is either a MDString or a MDNode with the first
1291 // operand a MDString.
1292 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1293 if (!MD || MD->getNumOperands() == 0)
1294 continue;
1295 S = dyn_cast<MDString>(MD->getOperand(0));
1296 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1297 Args.push_back(MD->getOperand(i));
1298 } else {
1299 S = dyn_cast<MDString>(LoopID->getOperand(i));
1300 assert(Args.size() == 0 && "too many arguments for MDString")((Args.size() == 0 && "too many arguments for MDString"
) ? static_cast<void> (0) : __assert_fail ("Args.size() == 0 && \"too many arguments for MDString\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1300, __PRETTY_FUNCTION__));
1301 }
1302
1303 if (!S)
1304 continue;
1305
1306 // Check if the hint starts with the loop metadata prefix.
1307 StringRef Name = S->getString();
1308 if (Args.size() == 1)
1309 setHint(Name, Args[0]);
1310 }
1311 }
1312
1313 /// Checks string hint with one operand and set value if valid.
1314 void setHint(StringRef Name, Metadata *Arg) {
1315 if (!Name.startswith(Prefix()))
1316 return;
1317 Name = Name.substr(Prefix().size(), StringRef::npos);
1318
1319 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1320 if (!C) return;
1321 unsigned Val = C->getZExtValue();
1322
1323 Hint *Hints[] = {&Width, &Interleave, &Force};
1324 for (auto H : Hints) {
1325 if (Name == H->Name) {
1326 if (H->validate(Val))
1327 H->Value = Val;
1328 else
1329 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: ignoring invalid hint '"
<< Name << "'\n"; } } while (0);
1330 break;
1331 }
1332 }
1333 }
1334
1335 /// Create a new hint from name / value pair.
1336 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1337 LLVMContext &Context = TheLoop->getHeader()->getContext();
1338 Metadata *MDs[] = {MDString::get(Context, Name),
1339 ConstantAsMetadata::get(
1340 ConstantInt::get(Type::getInt32Ty(Context), V))};
1341 return MDNode::get(Context, MDs);
1342 }
1343
1344 /// Matches metadata with hint name.
1345 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1346 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1347 if (!Name)
1348 return false;
1349
1350 for (auto H : HintTypes)
1351 if (Name->getString().endswith(H.Name))
1352 return true;
1353 return false;
1354 }
1355
1356 /// Sets current hints into loop metadata, keeping other values intact.
1357 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1358 if (HintTypes.size() == 0)
1359 return;
1360
1361 // Reserve the first element to LoopID (see below).
1362 SmallVector<Metadata *, 4> MDs(1);
1363 // If the loop already has metadata, then ignore the existing operands.
1364 MDNode *LoopID = TheLoop->getLoopID();
1365 if (LoopID) {
1366 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1367 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1368 // If node in update list, ignore old value.
1369 if (!matchesHintMetadataName(Node, HintTypes))
1370 MDs.push_back(Node);
1371 }
1372 }
1373
1374 // Now, add the missing hints.
1375 for (auto H : HintTypes)
1376 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1377
1378 // Replace current metadata node with new one.
1379 LLVMContext &Context = TheLoop->getHeader()->getContext();
1380 MDNode *NewLoopID = MDNode::get(Context, MDs);
1381 // Set operand 0 to refer to the loop id itself.
1382 NewLoopID->replaceOperandWith(0, NewLoopID);
1383
1384 TheLoop->setLoopID(NewLoopID);
1385 }
1386
1387 /// The loop these hints belong to.
1388 const Loop *TheLoop;
1389};
1390
1391static void emitMissedWarning(Function *F, Loop *L,
1392 const LoopVectorizeHints &LH) {
1393 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE"loop-vectorize", *F,
1394 L->getStartLoc(), LH.emitRemark());
1395
1396 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1397 if (LH.getWidth() != 1)
1398 emitLoopVectorizeWarning(
1399 F->getContext(), *F, L->getStartLoc(),
1400 "failed explicitly specified loop vectorization");
1401 else if (LH.getInterleave() != 1)
1402 emitLoopInterleaveWarning(
1403 F->getContext(), *F, L->getStartLoc(),
1404 "failed explicitly specified loop interleaving");
1405 }
1406}
1407
1408static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1409 if (L.empty())
1410 return V.push_back(&L);
1411
1412 for (Loop *InnerL : L)
1413 addInnerLoop(*InnerL, V);
1414}
1415
1416/// The LoopVectorize Pass.
1417struct LoopVectorize : public FunctionPass {
1418 /// Pass identification, replacement for typeid
1419 static char ID;
1420
1421 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1422 : FunctionPass(ID),
1423 DisableUnrolling(NoUnrolling),
1424 AlwaysVectorize(AlwaysVectorize) {
1425 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1426 }
1427
1428 ScalarEvolution *SE;
1429 LoopInfo *LI;
1430 TargetTransformInfo *TTI;
1431 DominatorTree *DT;
1432 BlockFrequencyInfo *BFI;
1433 TargetLibraryInfo *TLI;
1434 AliasAnalysis *AA;
1435 AssumptionCache *AC;
1436 LoopAccessAnalysis *LAA;
1437 bool DisableUnrolling;
1438 bool AlwaysVectorize;
1439
1440 BlockFrequency ColdEntryFreq;
1441
1442 bool runOnFunction(Function &F) override {
1443 SE = &getAnalysis<ScalarEvolution>();
1444 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1445 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1446 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1447 BFI = &getAnalysis<BlockFrequencyInfo>();
1448 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1449 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1450 AA = &getAnalysis<AliasAnalysis>();
1451 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1452 LAA = &getAnalysis<LoopAccessAnalysis>();
1453
1454 // Compute some weights outside of the loop over the loops. Compute this
1455 // using a BranchProbability to re-use its scaling math.
1456 const BranchProbability ColdProb(1, 5); // 20%
1457 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1458
1459 // If the target claims to have no vector registers don't attempt
1460 // vectorization.
1461 if (!TTI->getNumberOfRegisters(true))
1462 return false;
1463
1464 // Build up a worklist of inner-loops to vectorize. This is necessary as
1465 // the act of vectorizing or partially unrolling a loop creates new loops
1466 // and can invalidate iterators across the loops.
1467 SmallVector<Loop *, 8> Worklist;
1468
1469 for (Loop *L : *LI)
1470 addInnerLoop(*L, Worklist);
1471
1472 LoopsAnalyzed += Worklist.size();
1473
1474 // Now walk the identified inner loops.
1475 bool Changed = false;
1476 while (!Worklist.empty())
1477 Changed |= processLoop(Worklist.pop_back_val());
1478
1479 // Process each loop nest in the function.
1480 return Changed;
1481 }
1482
1483 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1484 SmallVector<Metadata *, 4> MDs;
1485 // Reserve first location for self reference to the LoopID metadata node.
1486 MDs.push_back(nullptr);
1487 bool IsUnrollMetadata = false;
1488 MDNode *LoopID = L->getLoopID();
1489 if (LoopID) {
1490 // First find existing loop unrolling disable metadata.
1491 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1492 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1493 if (MD) {
1494 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1495 IsUnrollMetadata =
1496 S && S->getString().startswith("llvm.loop.unroll.disable");
1497 }
1498 MDs.push_back(LoopID->getOperand(i));
1499 }
1500 }
1501
1502 if (!IsUnrollMetadata) {
1503 // Add runtime unroll disable metadata.
1504 LLVMContext &Context = L->getHeader()->getContext();
1505 SmallVector<Metadata *, 1> DisableOperands;
1506 DisableOperands.push_back(
1507 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1508 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1509 MDs.push_back(DisableNode);
1510 MDNode *NewLoopID = MDNode::get(Context, MDs);
1511 // Set operand 0 to refer to the loop id itself.
1512 NewLoopID->replaceOperandWith(0, NewLoopID);
1513 L->setLoopID(NewLoopID);
1514 }
1515 }
1516
1517 bool processLoop(Loop *L) {
1518 assert(L->empty() && "Only process inner loops.")((L->empty() && "Only process inner loops.") ? static_cast
<void> (0) : __assert_fail ("L->empty() && \"Only process inner loops.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1518, __PRETTY_FUNCTION__));
1519
1520#ifndef NDEBUG
1521 const std::string DebugLocStr = getDebugLocString(L);
1522#endif /* NDEBUG */
1523
1524 DEBUG(dbgs() << "\nLV: Checking a loop in \""do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in \""
<< L->getHeader()->getParent()->getName() <<
"\" from " << DebugLocStr << "\n"; } } while (0)
1525 << L->getHeader()->getParent()->getName() << "\" from "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in \""
<< L->getHeader()->getParent()->getName() <<
"\" from " << DebugLocStr << "\n"; } } while (0)
1526 << DebugLocStr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in \""
<< L->getHeader()->getParent()->getName() <<
"\" from " << DebugLocStr << "\n"; } } while (0);
1527
1528 LoopVectorizeHints Hints(L, DisableUnrolling);
1529
1530 DEBUG(dbgs() << "LV: Loop hints:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1531 << " force="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1532 << (Hints.getForce() == LoopVectorizeHints::FK_Disableddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1533 ? "disabled"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1534 : (Hints.getForce() == LoopVectorizeHints::FK_Enableddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1535 ? "enabled"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1536 : "?")) << " width=" << Hints.getWidth()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0)
1537 << " unroll=" << Hints.getInterleave() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop hints:" <<
" force=" << (Hints.getForce() == LoopVectorizeHints::
FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints
::FK_Enabled ? "enabled" : "?")) << " width=" << Hints
.getWidth() << " unroll=" << Hints.getInterleave(
) << "\n"; } } while (0);
1538
1539 // Function containing loop
1540 Function *F = L->getHeader()->getParent();
1541
1542 // Looking at the diagnostic output is the only way to determine if a loop
1543 // was vectorized (other than looking at the IR or machine code), so it
1544 // is important to generate an optimization remark for each loop. Most of
1545 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1546 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1547 // less verbose reporting vectorized loops and unvectorized loops that may
1548 // benefit from vectorization, respectively.
1549
1550 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1551 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n"
; } } while (0);
1552 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE"loop-vectorize", *F,
1553 L->getStartLoc(), Hints.emitRemark());
1554 return false;
1555 }
1556
1557 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1558 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n"
; } } while (0);
1559 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE"loop-vectorize", *F,
1560 L->getStartLoc(), Hints.emitRemark());
1561 return false;
1562 }
1563
1564 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1565 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n"
; } } while (0);
1566 emitOptimizationRemarkAnalysis(
1567 F->getContext(), DEBUG_TYPE"loop-vectorize", *F, L->getStartLoc(),
1568 "loop not vectorized: vector width and interleave count are "
1569 "explicitly set to 1");
1570 return false;
1571 }
1572
1573 // Check the loop for a trip count threshold:
1574 // do not vectorize loops with a tiny trip count.
1575 const unsigned TC = SE->getSmallConstantTripCount(L);
1576 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1577 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. "
<< "This loop is not worth vectorizing."; } } while (0
)
1578 << "This loop is not worth vectorizing.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. "
<< "This loop is not worth vectorizing."; } } while (0
);
1579 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1580 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << " But vectorizing was explicitly forced.\n"
; } } while (0);
1581 else {
1582 DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "\n"; } } while (0);
1583 emitOptimizationRemarkAnalysis(
1584 F->getContext(), DEBUG_TYPE"loop-vectorize", *F, L->getStartLoc(),
1585 "vectorization is not beneficial and is not explicitly forced");
1586 return false;
1587 }
1588 }
1589
1590 // Check if it is legal to vectorize the loop.
1591 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA);
1592 if (!LVL.canVectorize()) {
1593 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Cannot prove legality.\n"
; } } while (0);
1594 emitMissedWarning(F, L, Hints);
1595 return false;
1596 }
1597
1598 // Use the cost model.
1599 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1600
1601 // Check the function attributes to find out if this function should be
1602 // optimized for size.
1603 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1604 F->hasFnAttribute(Attribute::OptimizeForSize);
1605
1606 // Compute the weighted frequency of this loop being executed and see if it
1607 // is less than 20% of the function entry baseline frequency. Note that we
1608 // always have a canonical loop here because we think we *can* vectoriez.
1609 // FIXME: This is hidden behind a flag due to pervasive problems with
1610 // exactly what block frequency models.
1611 if (LoopVectorizeWithBlockFrequency) {
1612 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1613 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1614 LoopEntryFreq < ColdEntryFreq)
1615 OptForSize = true;
1616 }
1617
1618 // Check the function attributes to see if implicit floats are allowed.a
1619 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1620 // an integer loop and the vector instructions selected are purely integer
1621 // vector instructions?
1622 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1623 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
"attribute is used.\n"; } } while (0)
1624 "attribute is used.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
"attribute is used.\n"; } } while (0);
1625 emitOptimizationRemarkAnalysis(
1626 F->getContext(), DEBUG_TYPE"loop-vectorize", *F, L->getStartLoc(),
1627 "loop not vectorized due to NoImplicitFloat attribute");
1628 emitMissedWarning(F, L, Hints);
1629 return false;
1630 }
1631
1632 // Select the optimal vectorization factor.
1633 const LoopVectorizationCostModel::VectorizationFactor VF =
1634 CM.selectVectorizationFactor(OptForSize);
1635
1636 // Select the unroll factor.
1637 const unsigned UF =
1638 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1639
1640 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop ("
<< VF.Width << ") in " << DebugLocStr <<
'\n'; } } while (0)
1641 << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop ("
<< VF.Width << ") in " << DebugLocStr <<
'\n'; } } while (0);
1642 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Unroll Factor is "
<< UF << '\n'; } } while (0);
1643
1644 if (VF.Width == 1) {
1645 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Vectorization is possible but not beneficial\n"
; } } while (0);
1646
1647 if (UF == 1) {
1648 emitOptimizationRemarkAnalysis(
1649 F->getContext(), DEBUG_TYPE"loop-vectorize", *F, L->getStartLoc(),
1650 "not beneficial to vectorize and user disabled interleaving");
1651 return false;
1652 }
1653 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Trying to at least unroll the loops.\n"
; } } while (0);
1654
1655 // Report the unrolling decision.
1656 emitOptimizationRemark(F->getContext(), DEBUG_TYPE"loop-vectorize", *F, L->getStartLoc(),
1657 Twine("unrolled with interleaving factor " +
1658 Twine(UF) +
1659 " (vectorization not beneficial)"));
1660
1661 // We decided not to vectorize, but we may want to unroll.
1662
1663 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, UF);
1664 Unroller.vectorize(&LVL);
1665 } else {
1666 // If we decided that it is *legal* to vectorize the loop then do it.
1667 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, UF);
1668 LB.vectorize(&LVL);
1669 ++LoopsVectorized;
1670
1671 // Add metadata to disable runtime unrolling scalar loop when there's no
1672 // runtime check about strides and memory. Because at this situation,
1673 // scalar loop is rarely used not worthy to be unrolled.
1674 if (!LB.IsSafetyChecksAdded())
1675 AddRuntimeUnrollDisableMetaData(L);
1676
1677 // Report the vectorization decision.
1678 emitOptimizationRemark(
1679 F->getContext(), DEBUG_TYPE"loop-vectorize", *F, L->getStartLoc(),
1680 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1681 ", unrolling interleave factor: " + Twine(UF) + ")");
1682 }
1683
1684 // Mark the loop as already vectorized to avoid vectorizing again.
1685 Hints.setAlreadyVectorized();
1686
1687 DEBUG(verifyFunction(*L->getHeader()->getParent()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { verifyFunction(*L->getHeader()->getParent
()); } } while (0);
1688 return true;
1689 }
1690
1691 void getAnalysisUsage(AnalysisUsage &AU) const override {
1692 AU.addRequired<AssumptionCacheTracker>();
1693 AU.addRequiredID(LoopSimplifyID);
1694 AU.addRequiredID(LCSSAID);
1695 AU.addRequired<BlockFrequencyInfo>();
1696 AU.addRequired<DominatorTreeWrapperPass>();
1697 AU.addRequired<LoopInfoWrapperPass>();
1698 AU.addRequired<ScalarEvolution>();
1699 AU.addRequired<TargetTransformInfoWrapperPass>();
1700 AU.addRequired<AliasAnalysis>();
1701 AU.addRequired<LoopAccessAnalysis>();
1702 AU.addPreserved<LoopInfoWrapperPass>();
1703 AU.addPreserved<DominatorTreeWrapperPass>();
1704 AU.addPreserved<AliasAnalysis>();
1705 }
1706
1707};
1708
1709} // end anonymous namespace
1710
1711//===----------------------------------------------------------------------===//
1712// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1713// LoopVectorizationCostModel.
1714//===----------------------------------------------------------------------===//
1715
1716Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1717 // We need to place the broadcast of invariant variables outside the loop.
1718 Instruction *Instr = dyn_cast<Instruction>(V);
1719 bool NewInstr =
1720 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1721 Instr->getParent()) != LoopVectorBody.end());
1722 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1723
1724 // Place the code for broadcasting invariant variables in the new preheader.
1725 IRBuilder<>::InsertPointGuard Guard(Builder);
1726 if (Invariant)
1727 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1728
1729 // Broadcast the scalar into all locations in the vector.
1730 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1731
1732 return Shuf;
1733}
1734
1735Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1736 Value *Step) {
1737 assert(Val->getType()->isVectorTy() && "Must be a vector")((Val->getType()->isVectorTy() && "Must be a vector"
) ? static_cast<void> (0) : __assert_fail ("Val->getType()->isVectorTy() && \"Must be a vector\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1737, __PRETTY_FUNCTION__));
1738 assert(Val->getType()->getScalarType()->isIntegerTy() &&((Val->getType()->getScalarType()->isIntegerTy() &&
"Elem must be an integer") ? static_cast<void> (0) : __assert_fail
("Val->getType()->getScalarType()->isIntegerTy() && \"Elem must be an integer\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1739, __PRETTY_FUNCTION__))
1739 "Elem must be an integer")((Val->getType()->getScalarType()->isIntegerTy() &&
"Elem must be an integer") ? static_cast<void> (0) : __assert_fail
("Val->getType()->getScalarType()->isIntegerTy() && \"Elem must be an integer\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1739, __PRETTY_FUNCTION__));
1740 assert(Step->getType() == Val->getType()->getScalarType() &&((Step->getType() == Val->getType()->getScalarType()
&& "Step has wrong type") ? static_cast<void> (
0) : __assert_fail ("Step->getType() == Val->getType()->getScalarType() && \"Step has wrong type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1741, __PRETTY_FUNCTION__))
1741 "Step has wrong type")((Step->getType() == Val->getType()->getScalarType()
&& "Step has wrong type") ? static_cast<void> (
0) : __assert_fail ("Step->getType() == Val->getType()->getScalarType() && \"Step has wrong type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1741, __PRETTY_FUNCTION__));
1742 // Create the types.
1743 Type *ITy = Val->getType()->getScalarType();
1744 VectorType *Ty = cast<VectorType>(Val->getType());
1745 int VLen = Ty->getNumElements();
1746 SmallVector<Constant*, 8> Indices;
1747
1748 // Create a vector of consecutive numbers from zero to VF.
1749 for (int i = 0; i < VLen; ++i)
1750 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1751
1752 // Add the consecutive indices to the vector value.
1753 Constant *Cv = ConstantVector::get(Indices);
1754 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec")((Cv->getType() == Val->getType() && "Invalid consecutive vec"
) ? static_cast<void> (0) : __assert_fail ("Cv->getType() == Val->getType() && \"Invalid consecutive vec\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1754, __PRETTY_FUNCTION__));
1755 Step = Builder.CreateVectorSplat(VLen, Step);
1756 assert(Step->getType() == Val->getType() && "Invalid step vec")((Step->getType() == Val->getType() && "Invalid step vec"
) ? static_cast<void> (0) : __assert_fail ("Step->getType() == Val->getType() && \"Invalid step vec\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1756, __PRETTY_FUNCTION__));
1757 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1758 // which can be found from the original scalar operations.
1759 Step = Builder.CreateMul(Cv, Step);
1760 return Builder.CreateAdd(Val, Step, "induction");
1761}
1762
1763/// \brief Find the operand of the GEP that should be checked for consecutive
1764/// stores. This ignores trailing indices that have no effect on the final
1765/// pointer.
1766static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep) {
1767 const DataLayout &DL = Gep->getModule()->getDataLayout();
1768 unsigned LastOperand = Gep->getNumOperands() - 1;
1769 unsigned GEPAllocSize = DL.getTypeAllocSize(
1770 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1771
1772 // Walk backwards and try to peel off zeros.
1773 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1774 // Find the type we're currently indexing into.
1775 gep_type_iterator GEPTI = gep_type_begin(Gep);
1776 std::advance(GEPTI, LastOperand - 1);
1777
1778 // If it's a type with the same allocation size as the result of the GEP we
1779 // can peel off the zero index.
1780 if (DL.getTypeAllocSize(*GEPTI) != GEPAllocSize)
1781 break;
1782 --LastOperand;
1783 }
1784
1785 return LastOperand;
1786}
1787
1788int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1789 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr")((Ptr->getType()->isPointerTy() && "Unexpected non-ptr"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->isPointerTy() && \"Unexpected non-ptr\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1789, __PRETTY_FUNCTION__));
1790 // Make sure that the pointer does not point to structs.
1791 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1792 return 0;
1793
1794 // If this value is a pointer induction variable we know it is consecutive.
1795 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1796 if (Phi && Inductions.count(Phi)) {
1797 InductionInfo II = Inductions[Phi];
1798 return II.getConsecutiveDirection();
1799 }
1800
1801 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1802 if (!Gep)
1803 return 0;
1804
1805 unsigned NumOperands = Gep->getNumOperands();
1806 Value *GpPtr = Gep->getPointerOperand();
1807 // If this GEP value is a consecutive pointer induction variable and all of
1808 // the indices are constant then we know it is consecutive. We can
1809 Phi = dyn_cast<PHINode>(GpPtr);
1810 if (Phi && Inductions.count(Phi)) {
1811
1812 // Make sure that the pointer does not point to structs.
1813 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1814 if (GepPtrType->getElementType()->isAggregateType())
1815 return 0;
1816
1817 // Make sure that all of the index operands are loop invariant.
1818 for (unsigned i = 1; i < NumOperands; ++i)
1819 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1820 return 0;
1821
1822 InductionInfo II = Inductions[Phi];
1823 return II.getConsecutiveDirection();
1824 }
1825
1826 unsigned InductionOperand = getGEPInductionOperand(Gep);
1827
1828 // Check that all of the gep indices are uniform except for our induction
1829 // operand.
1830 for (unsigned i = 0; i != NumOperands; ++i)
1831 if (i != InductionOperand &&
1832 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1833 return 0;
1834
1835 // We can emit wide load/stores only if the last non-zero index is the
1836 // induction variable.
1837 const SCEV *Last = nullptr;
1838 if (!Strides.count(Gep))
1839 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1840 else {
1841 // Because of the multiplication by a stride we can have a s/zext cast.
1842 // We are going to replace this stride by 1 so the cast is safe to ignore.
1843 //
1844 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1845 // %0 = trunc i64 %indvars.iv to i32
1846 // %mul = mul i32 %0, %Stride1
1847 // %idxprom = zext i32 %mul to i64 << Safe cast.
1848 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1849 //
1850 Last = replaceSymbolicStrideSCEV(SE, Strides,
1851 Gep->getOperand(InductionOperand), Gep);
1852 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1853 Last =
1854 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1855 ? C->getOperand()
1856 : Last;
1857 }
1858 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1859 const SCEV *Step = AR->getStepRecurrence(*SE);
1860
1861 // The memory is consecutive because the last index is consecutive
1862 // and all other indices are loop invariant.
1863 if (Step->isOne())
1864 return 1;
1865 if (Step->isAllOnesValue())
1866 return -1;
1867 }
1868
1869 return 0;
1870}
1871
1872bool LoopVectorizationLegality::isUniform(Value *V) {
1873 return LAI->isUniform(V);
1874}
1875
1876InnerLoopVectorizer::VectorParts&
1877InnerLoopVectorizer::getVectorValue(Value *V) {
1878 assert(V != Induction && "The new induction variable should not be used.")((V != Induction && "The new induction variable should not be used."
) ? static_cast<void> (0) : __assert_fail ("V != Induction && \"The new induction variable should not be used.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1878, __PRETTY_FUNCTION__));
1879 assert(!V->getType()->isVectorTy() && "Can't widen a vector")((!V->getType()->isVectorTy() && "Can't widen a vector"
) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isVectorTy() && \"Can't widen a vector\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1879, __PRETTY_FUNCTION__));
1880
1881 // If we have a stride that is replaced by one, do it here.
1882 if (Legal->hasStride(V))
1883 V = ConstantInt::get(V->getType(), 1);
1884
1885 // If we have this scalar in the map, return it.
1886 if (WidenMap.has(V))
1887 return WidenMap.get(V);
1888
1889 // If this scalar is unknown, assume that it is a constant or that it is
1890 // loop invariant. Broadcast V and save the value for future uses.
1891 Value *B = getBroadcastInstrs(V);
1892 return WidenMap.splat(V, B);
1893}
1894
1895Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1896 assert(Vec->getType()->isVectorTy() && "Invalid type")((Vec->getType()->isVectorTy() && "Invalid type"
) ? static_cast<void> (0) : __assert_fail ("Vec->getType()->isVectorTy() && \"Invalid type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1896, __PRETTY_FUNCTION__));
1897 SmallVector<Constant*, 8> ShuffleMask;
1898 for (unsigned i = 0; i < VF; ++i)
1899 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1900
1901 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1902 ConstantVector::get(ShuffleMask),
1903 "reverse");
1904}
1905
1906// Get a mask to interleave \p NumVec vectors into a wide vector.
1907// I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
1908// E.g. For 2 interleaved vectors, if VF is 4, the mask is:
1909// <0, 4, 1, 5, 2, 6, 3, 7>
1910static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
1911 unsigned NumVec) {
1912 SmallVector<Constant *, 16> Mask;
1913 for (unsigned i = 0; i < VF; i++)
1914 for (unsigned j = 0; j < NumVec; j++)
1915 Mask.push_back(Builder.getInt32(j * VF + i));
1916
1917 return ConstantVector::get(Mask);
1918}
1919
1920// Get the strided mask starting from index \p Start.
1921// I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
1922static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
1923 unsigned Stride, unsigned VF) {
1924 SmallVector<Constant *, 16> Mask;
1925 for (unsigned i = 0; i < VF; i++)
1926 Mask.push_back(Builder.getInt32(Start + i * Stride));
1927
1928 return ConstantVector::get(Mask);
1929}
1930
1931// Get a mask of two parts: The first part consists of sequential integers
1932// starting from 0, The second part consists of UNDEFs.
1933// I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
1934static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
1935 unsigned NumUndef) {
1936 SmallVector<Constant *, 16> Mask;
1937 for (unsigned i = 0; i < NumInt; i++)
1938 Mask.push_back(Builder.getInt32(i));
1939
1940 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
1941 for (unsigned i = 0; i < NumUndef; i++)
1942 Mask.push_back(Undef);
1943
1944 return ConstantVector::get(Mask);
1945}
1946
1947// Concatenate two vectors with the same element type. The 2nd vector should
1948// not have more elements than the 1st vector. If the 2nd vector has less
1949// elements, extend it with UNDEFs.
1950static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
1951 Value *V2) {
1952 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
1953 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
1954 assert(VecTy1 && VecTy2 &&((VecTy1 && VecTy2 && VecTy1->getScalarType
() == VecTy2->getScalarType() && "Expect two vectors with the same element type"
) ? static_cast<void> (0) : __assert_fail ("VecTy1 && VecTy2 && VecTy1->getScalarType() == VecTy2->getScalarType() && \"Expect two vectors with the same element type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1956, __PRETTY_FUNCTION__))
1955 VecTy1->getScalarType() == VecTy2->getScalarType() &&((VecTy1 && VecTy2 && VecTy1->getScalarType
() == VecTy2->getScalarType() && "Expect two vectors with the same element type"
) ? static_cast<void> (0) : __assert_fail ("VecTy1 && VecTy2 && VecTy1->getScalarType() == VecTy2->getScalarType() && \"Expect two vectors with the same element type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1956, __PRETTY_FUNCTION__))
1956 "Expect two vectors with the same element type")((VecTy1 && VecTy2 && VecTy1->getScalarType
() == VecTy2->getScalarType() && "Expect two vectors with the same element type"
) ? static_cast<void> (0) : __assert_fail ("VecTy1 && VecTy2 && VecTy1->getScalarType() == VecTy2->getScalarType() && \"Expect two vectors with the same element type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1956, __PRETTY_FUNCTION__));
1957
1958 unsigned NumElts1 = VecTy1->getNumElements();
1959 unsigned NumElts2 = VecTy2->getNumElements();
1960 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements")((NumElts1 >= NumElts2 && "Unexpect the first vector has less elements"
) ? static_cast<void> (0) : __assert_fail ("NumElts1 >= NumElts2 && \"Unexpect the first vector has less elements\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1960, __PRETTY_FUNCTION__));
1961
1962 if (NumElts1 > NumElts2) {
1963 // Extend with UNDEFs.
1964 Constant *ExtMask =
1965 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
1966 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
1967 }
1968
1969 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
1970 return Builder.CreateShuffleVector(V1, V2, Mask);
1971}
1972
1973// Concatenate vectors in the given list. All vectors have the same type.
1974static Value *ConcatenateVectors(IRBuilder<> &Builder,
1975 ArrayRef<Value *> InputList) {
1976 unsigned NumVec = InputList.size();
1977 assert(NumVec > 1 && "Should be at least two vectors")((NumVec > 1 && "Should be at least two vectors") ?
static_cast<void> (0) : __assert_fail ("NumVec > 1 && \"Should be at least two vectors\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1977, __PRETTY_FUNCTION__));
1978
1979 SmallVector<Value *, 8> ResList;
1980 ResList.append(InputList.begin(), InputList.end());
1981 do {
1982 SmallVector<Value *, 8> TmpList;
1983 for (unsigned i = 0; i < NumVec - 1; i += 2) {
1984 Value *V0 = ResList[i], *V1 = ResList[i + 1];
1985 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&(((V0->getType() == V1->getType() || i == NumVec - 2) &&
"Only the last vector may have a different type") ? static_cast
<void> (0) : __assert_fail ("(V0->getType() == V1->getType() || i == NumVec - 2) && \"Only the last vector may have a different type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1986, __PRETTY_FUNCTION__))
1986 "Only the last vector may have a different type")(((V0->getType() == V1->getType() || i == NumVec - 2) &&
"Only the last vector may have a different type") ? static_cast
<void> (0) : __assert_fail ("(V0->getType() == V1->getType() || i == NumVec - 2) && \"Only the last vector may have a different type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1986, __PRETTY_FUNCTION__));
1987
1988 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
1989 }
1990
1991 // Push the last vector if the total number of vectors is odd.
1992 if (NumVec % 2 != 0)
1993 TmpList.push_back(ResList[NumVec - 1]);
1994
1995 ResList = TmpList;
1996 NumVec = ResList.size();
1997 } while (NumVec > 1);
1998
1999 return ResList[0];
2000}
2001
2002// Try to vectorize the interleave group that \p Instr belongs to.
2003//
2004// E.g. Translate following interleaved load group (factor = 3):
2005// for (i = 0; i < N; i+=3) {
2006// R = Pic[i]; // Member of index 0
2007// G = Pic[i+1]; // Member of index 1
2008// B = Pic[i+2]; // Member of index 2
2009// ... // do something to R, G, B
2010// }
2011// To:
2012// %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2013// %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2014// %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2015// %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2016//
2017// Or translate following interleaved store group (factor = 3):
2018// for (i = 0; i < N; i+=3) {
2019// ... do something to R, G, B
2020// Pic[i] = R; // Member of index 0
2021// Pic[i+1] = G; // Member of index 1
2022// Pic[i+2] = B; // Member of index 2
2023// }
2024// To:
2025// %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2026// %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2027// %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2028// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2029// store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2030void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2031 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2032 assert(Group && "Fail to get an interleaved access group.")((Group && "Fail to get an interleaved access group."
) ? static_cast<void> (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2032, __PRETTY_FUNCTION__));
2033
2034 // Skip if current instruction is not the insert position.
2035 if (Instr != Group->getInsertPos())
2036 return;
2037
2038 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2039 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2040 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2041
2042 // Prepare for the vector type of the interleaved load/store.
2043 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2044 unsigned InterleaveFactor = Group->getFactor();
2045 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2046 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2047
2048 // Prepare for the new pointers.
2049 setDebugLocFromInst(Builder, Ptr);
2050 VectorParts &PtrParts = getVectorValue(Ptr);
2051 SmallVector<Value *, 2> NewPtrs;
2052 unsigned Index = Group->getIndex(Instr);
2053 for (unsigned Part = 0; Part < UF; Part++) {
2054 // Extract the pointer for current instruction from the pointer vector. A
2055 // reverse access uses the pointer in the last lane.
2056 Value *NewPtr = Builder.CreateExtractElement(
2057 PtrParts[Part],
2058 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2059
2060 // Notice current instruction could be any index. Need to adjust the address
2061 // to the member of index 0.
2062 //
2063 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2064 // b = A[i]; // Member of index 0
2065 // Current pointer is pointed to A[i+1], adjust it to A[i].
2066 //
2067 // E.g. A[i+1] = a; // Member of index 1
2068 // A[i] = b; // Member of index 0
2069 // A[i+2] = c; // Member of index 2 (Current instruction)
2070 // Current pointer is pointed to A[i+2], adjust it to A[i].
2071 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2072
2073 // Cast to the vector pointer type.
2074 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2075 }
2076
2077 setDebugLocFromInst(Builder, Instr);
2078 Value *UndefVec = UndefValue::get(VecTy);
2079
2080 // Vectorize the interleaved load group.
2081 if (LI) {
2082 for (unsigned Part = 0; Part < UF; Part++) {
2083 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2084 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2085
2086 for (unsigned i = 0; i < InterleaveFactor; i++) {
2087 Instruction *Member = Group->getMember(i);
2088
2089 // Skip the gaps in the group.
2090 if (!Member)
2091 continue;
2092
2093 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2094 Value *StridedVec = Builder.CreateShuffleVector(
2095 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2096
2097 // If this member has different type, cast the result type.
2098 if (Member->getType() != ScalarTy) {
2099 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2100 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2101 }
2102
2103 VectorParts &Entry = WidenMap.get(Member);
2104 Entry[Part] =
2105 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2106 }
2107
2108 propagateMetadata(NewLoadInstr, Instr);
2109 }
2110 return;
2111 }
2112
2113 // The sub vector type for current instruction.
2114 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2115
2116 // Vectorize the interleaved store group.
2117 for (unsigned Part = 0; Part < UF; Part++) {
2118 // Collect the stored vector from each member.
2119 SmallVector<Value *, 4> StoredVecs;
2120 for (unsigned i = 0; i < InterleaveFactor; i++) {
2121 // Interleaved store group doesn't allow a gap, so each index has a member
2122 Instruction *Member = Group->getMember(i);
2123 assert(Member && "Fail to get a member from an interleaved store group")((Member && "Fail to get a member from an interleaved store group"
) ? static_cast<void> (0) : __assert_fail ("Member && \"Fail to get a member from an interleaved store group\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2123, __PRETTY_FUNCTION__));
2124
2125 Value *StoredVec =
2126 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2127 if (Group->isReverse())
2128 StoredVec = reverseVector(StoredVec);
2129
2130 // If this member has different type, cast it to an unified type.
2131 if (StoredVec->getType() != SubVT)
2132 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2133
2134 StoredVecs.push_back(StoredVec);
2135 }
2136
2137 // Concatenate all vectors into a wide vector.
2138 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2139
2140 // Interleave the elements in the wide vector.
2141 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2142 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2143 "interleaved.vec");
2144
2145 Instruction *NewStoreInstr =
2146 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2147 propagateMetadata(NewStoreInstr, Instr);
2148 }
2149}
2150
2151void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2152 // Attempt to issue a wide load.
2153 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2154 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2155
2156 assert((LI || SI) && "Invalid Load/Store instruction")(((LI || SI) && "Invalid Load/Store instruction") ? static_cast
<void> (0) : __assert_fail ("(LI || SI) && \"Invalid Load/Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2156, __PRETTY_FUNCTION__));
2157
2158 // Try to vectorize the interleave group if this access is interleaved.
2159 if (Legal->isAccessInterleaved(Instr))
2160 return vectorizeInterleaveGroup(Instr);
2161
2162 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2163 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2164 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2165 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2166 // An alignment of 0 means target abi alignment. We need to use the scalar's
2167 // target abi alignment in such a case.
2168 const DataLayout &DL = Instr->getModule()->getDataLayout();
2169 if (!Alignment)
2170 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2171 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2172 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2173 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2174
2175 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2176 !Legal->isMaskRequired(SI))
2177 return scalarizeInstruction(Instr, true);
2178
2179 if (ScalarAllocatedSize != VectorElementSize)
2180 return scalarizeInstruction(Instr);
2181
2182 // If the pointer is loop invariant or if it is non-consecutive,
2183 // scalarize the load.
2184 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2185 bool Reverse = ConsecutiveStride < 0;
2186 bool UniformLoad = LI && Legal->isUniform(Ptr);
2187 if (!ConsecutiveStride || UniformLoad)
2188 return scalarizeInstruction(Instr);
2189
2190 Constant *Zero = Builder.getInt32(0);
2191 VectorParts &Entry = WidenMap.get(Instr);
2192
2193 // Handle consecutive loads/stores.
2194 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2195 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2196 setDebugLocFromInst(Builder, Gep);
2197 Value *PtrOperand = Gep->getPointerOperand();
2198 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2199 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2200
2201 // Create the new GEP with the new induction variable.
2202 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2203 Gep2->setOperand(0, FirstBasePtr);
2204 Gep2->setName("gep.indvar.base");
2205 Ptr = Builder.Insert(Gep2);
2206 } else if (Gep) {
2207 setDebugLocFromInst(Builder, Gep);
2208 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),((SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand
()), OrigLoop) && "Base ptr must be invariant") ? static_cast
<void> (0) : __assert_fail ("SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), OrigLoop) && \"Base ptr must be invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2209, __PRETTY_FUNCTION__))
2209 OrigLoop) && "Base ptr must be invariant")((SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand
()), OrigLoop) && "Base ptr must be invariant") ? static_cast
<void> (0) : __assert_fail ("SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), OrigLoop) && \"Base ptr must be invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2209, __PRETTY_FUNCTION__));
2210
2211 // The last index does not have to be the induction. It can be
2212 // consecutive and be a function of the index. For example A[I+1];
2213 unsigned NumOperands = Gep->getNumOperands();
2214 unsigned InductionOperand = getGEPInductionOperand(Gep);
2215 // Create the new GEP with the new induction variable.
2216 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2217
2218 for (unsigned i = 0; i < NumOperands; ++i) {
2219 Value *GepOperand = Gep->getOperand(i);
2220 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2221
2222 // Update last index or loop invariant instruction anchored in loop.
2223 if (i == InductionOperand ||
2224 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2225 assert((i == InductionOperand ||(((i == InductionOperand || SE->isLoopInvariant(SE->getSCEV
(GepOperandInst), OrigLoop)) && "Must be last index or loop invariant"
) ? static_cast<void> (0) : __assert_fail ("(i == InductionOperand || SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) && \"Must be last index or loop invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2227, __PRETTY_FUNCTION__))
2226 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&(((i == InductionOperand || SE->isLoopInvariant(SE->getSCEV
(GepOperandInst), OrigLoop)) && "Must be last index or loop invariant"
) ? static_cast<void> (0) : __assert_fail ("(i == InductionOperand || SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) && \"Must be last index or loop invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2227, __PRETTY_FUNCTION__))
2227 "Must be last index or loop invariant")(((i == InductionOperand || SE->isLoopInvariant(SE->getSCEV
(GepOperandInst), OrigLoop)) && "Must be last index or loop invariant"
) ? static_cast<void> (0) : __assert_fail ("(i == InductionOperand || SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) && \"Must be last index or loop invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2227, __PRETTY_FUNCTION__));
2228
2229 VectorParts &GEPParts = getVectorValue(GepOperand);
2230 Value *Index = GEPParts[0];
2231 Index = Builder.CreateExtractElement(Index, Zero);
2232 Gep2->setOperand(i, Index);
2233 Gep2->setName("gep.indvar.idx");
2234 }
2235 }
2236 Ptr = Builder.Insert(Gep2);
2237 } else {
2238 // Use the induction element ptr.
2239 assert(isa<PHINode>(Ptr) && "Invalid induction ptr")((isa<PHINode>(Ptr) && "Invalid induction ptr")
? static_cast<void> (0) : __assert_fail ("isa<PHINode>(Ptr) && \"Invalid induction ptr\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2239, __PRETTY_FUNCTION__));
2240 setDebugLocFromInst(Builder, Ptr);
2241 VectorParts &PtrVal = getVectorValue(Ptr);
2242 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2243 }
2244
2245 VectorParts Mask = createBlockInMask(Instr->getParent());
2246 // Handle Stores:
2247 if (SI) {
2248 assert(!Legal->isUniform(SI->getPointerOperand()) &&((!Legal->isUniform(SI->getPointerOperand()) &&
"We do not allow storing to uniform addresses") ? static_cast
<void> (0) : __assert_fail ("!Legal->isUniform(SI->getPointerOperand()) && \"We do not allow storing to uniform addresses\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2249, __PRETTY_FUNCTION__))
2249 "We do not allow storing to uniform addresses")((!Legal->isUniform(SI->getPointerOperand()) &&
"We do not allow storing to uniform addresses") ? static_cast
<void> (0) : __assert_fail ("!Legal->isUniform(SI->getPointerOperand()) && \"We do not allow storing to uniform addresses\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2249, __PRETTY_FUNCTION__));
2250 setDebugLocFromInst(Builder, SI);
2251 // We don't want to update the value in the map as it might be used in
2252 // another expression. So don't use a reference type for "StoredVal".
2253 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2254
2255 for (unsigned Part = 0; Part < UF; ++Part) {
2256 // Calculate the pointer for the specific unroll-part.
2257 Value *PartPtr =
2258 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2259
2260 if (Reverse) {
2261 // If we store to reverse consecutive memory locations then we need
2262 // to reverse the order of elements in the stored value.
2263 StoredVal[Part] = reverseVector(StoredVal[Part]);
2264 // If the address is consecutive but reversed, then the
2265 // wide store needs to start at the last vector element.
2266 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2267 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2268 Mask[Part] = reverseVector(Mask[Part]);
2269 }
2270
2271 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2272 DataTy->getPointerTo(AddressSpace));
2273
2274 Instruction *NewSI;
2275 if (Legal->isMaskRequired(SI))
2276 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2277 Mask[Part]);
2278 else
2279 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2280 propagateMetadata(NewSI, SI);
2281 }
2282 return;
2283 }
2284
2285 // Handle loads.
2286 assert(LI && "Must have a load instruction")((LI && "Must have a load instruction") ? static_cast
<void> (0) : __assert_fail ("LI && \"Must have a load instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2286, __PRETTY_FUNCTION__));
2287 setDebugLocFromInst(Builder, LI);
2288 for (unsigned Part = 0; Part < UF; ++Part) {
2289 // Calculate the pointer for the specific unroll-part.
2290 Value *PartPtr =
2291 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2292
2293 if (Reverse) {
2294 // If the address is consecutive but reversed, then the
2295 // wide load needs to start at the last vector element.
2296 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2297 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2298 Mask[Part] = reverseVector(Mask[Part]);
2299 }
2300
2301 Instruction* NewLI;
2302 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2303 DataTy->getPointerTo(AddressSpace));
2304 if (Legal->isMaskRequired(LI))
2305 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2306 UndefValue::get(DataTy),
2307 "wide.masked.load");
2308 else
2309 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2310 propagateMetadata(NewLI, LI);
2311 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2312 }
2313}
2314
2315void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2316 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors")((!Instr->getType()->isAggregateType() && "Can't handle vectors"
) ? static_cast<void> (0) : __assert_fail ("!Instr->getType()->isAggregateType() && \"Can't handle vectors\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2316, __PRETTY_FUNCTION__));
2317 // Holds vector parameters or scalars, in case of uniform vals.
2318 SmallVector<VectorParts, 4> Params;
2319
2320 setDebugLocFromInst(Builder, Instr);
2321
2322 // Find all of the vectorized parameters.
2323 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2324 Value *SrcOp = Instr->getOperand(op);
2325
2326 // If we are accessing the old induction variable, use the new one.
2327 if (SrcOp == OldInduction) {
2328 Params.push_back(getVectorValue(SrcOp));
2329 continue;
2330 }
2331
2332 // Try using previously calculated values.
2333 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2334
2335 // If the src is an instruction that appeared earlier in the basic block
2336 // then it should already be vectorized.
2337 if (SrcInst && OrigLoop->contains(SrcInst)) {
2338 assert(WidenMap.has(SrcInst) && "Source operand is unavailable")((WidenMap.has(SrcInst) && "Source operand is unavailable"
) ? static_cast<void> (0) : __assert_fail ("WidenMap.has(SrcInst) && \"Source operand is unavailable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2338, __PRETTY_FUNCTION__));
2339 // The parameter is a vector value from earlier.
2340 Params.push_back(WidenMap.get(SrcInst));
2341 } else {
2342 // The parameter is a scalar from outside the loop. Maybe even a constant.
2343 VectorParts Scalars;
2344 Scalars.append(UF, SrcOp);
2345 Params.push_back(Scalars);
2346 }
2347 }
2348
2349 assert(Params.size() == Instr->getNumOperands() &&((Params.size() == Instr->getNumOperands() && "Invalid number of operands"
) ? static_cast<void> (0) : __assert_fail ("Params.size() == Instr->getNumOperands() && \"Invalid number of operands\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2350, __PRETTY_FUNCTION__))
2350 "Invalid number of operands")((Params.size() == Instr->getNumOperands() && "Invalid number of operands"
) ? static_cast<void> (0) : __assert_fail ("Params.size() == Instr->getNumOperands() && \"Invalid number of operands\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2350, __PRETTY_FUNCTION__));
2351
2352 // Does this instruction return a value ?
2353 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2354
2355 Value *UndefVec = IsVoidRetTy ? nullptr :
2356 UndefValue::get(VectorType::get(Instr->getType(), VF));
2357 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2358 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2359
2360 Instruction *InsertPt = Builder.GetInsertPoint();
2361 BasicBlock *IfBlock = Builder.GetInsertBlock();
2362 BasicBlock *CondBlock = nullptr;
2363
2364 VectorParts Cond;
2365 Loop *VectorLp = nullptr;
2366 if (IfPredicateStore) {
2367 assert(Instr->getParent()->getSinglePredecessor() &&((Instr->getParent()->getSinglePredecessor() &&
"Only support single predecessor blocks") ? static_cast<void
> (0) : __assert_fail ("Instr->getParent()->getSinglePredecessor() && \"Only support single predecessor blocks\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2368, __PRETTY_FUNCTION__))
2368 "Only support single predecessor blocks")((Instr->getParent()->getSinglePredecessor() &&
"Only support single predecessor blocks") ? static_cast<void
> (0) : __assert_fail ("Instr->getParent()->getSinglePredecessor() && \"Only support single predecessor blocks\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2368, __PRETTY_FUNCTION__));
2369 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2370 Instr->getParent());
2371 VectorLp = LI->getLoopFor(IfBlock);
2372 assert(VectorLp && "Must have a loop for this block")((VectorLp && "Must have a loop for this block") ? static_cast
<void> (0) : __assert_fail ("VectorLp && \"Must have a loop for this block\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2372, __PRETTY_FUNCTION__));
2373 }
2374
2375 // For each vector unroll 'part':
2376 for (unsigned Part = 0; Part < UF; ++Part) {
2377 // For each scalar that we create:
2378 for (unsigned Width = 0; Width < VF; ++Width) {
2379
2380 // Start if-block.
2381 Value *Cmp = nullptr;
2382 if (IfPredicateStore) {
2383 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2384 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2385 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2386 LoopVectorBody.push_back(CondBlock);
2387 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2388 // Update Builder with newly created basic block.
2389 Builder.SetInsertPoint(InsertPt);
2390 }
2391
2392 Instruction *Cloned = Instr->clone();
2393 if (!IsVoidRetTy)
2394 Cloned->setName(Instr->getName() + ".cloned");
2395 // Replace the operands of the cloned instructions with extracted scalars.
2396 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2397 Value *Op = Params[op][Part];
2398 // Param is a vector. Need to extract the right lane.
2399 if (Op->getType()->isVectorTy())
2400 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2401 Cloned->setOperand(op, Op);
2402 }
2403
2404 // Place the cloned scalar in the new loop.
2405 Builder.Insert(Cloned);
2406
2407 // If the original scalar returns a value we need to place it in a vector
2408 // so that future users will be able to use it.
2409 if (!IsVoidRetTy)
2410 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2411 Builder.getInt32(Width));
2412 // End if-block.
2413 if (IfPredicateStore) {
2414 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2415 LoopVectorBody.push_back(NewIfBlock);
2416 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2417 Builder.SetInsertPoint(InsertPt);
2418 Instruction *OldBr = IfBlock->getTerminator();
2419 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2420 OldBr->eraseFromParent();
2421 IfBlock = NewIfBlock;
2422 }
2423 }
2424 }
2425}
2426
2427static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2428 Instruction *Loc) {
2429 if (FirstInst)
2430 return FirstInst;
2431 if (Instruction *I = dyn_cast<Instruction>(V))
2432 return I->getParent() == Loc->getParent() ? I : nullptr;
2433 return nullptr;
2434}
2435
2436std::pair<Instruction *, Instruction *>
2437InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2438 Instruction *tnullptr = nullptr;
2439 if (!Legal->mustCheckStrides())
2440 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2441
2442 IRBuilder<> ChkBuilder(Loc);
2443
2444 // Emit checks.
2445 Value *Check = nullptr;
2446 Instruction *FirstInst = nullptr;
2447 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2448 SE = Legal->strides_end();
2449 SI != SE; ++SI) {
2450 Value *Ptr = stripIntegerCast(*SI);
2451 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2452 "stride.chk");
2453 // Store the first instruction we create.
2454 FirstInst = getFirstInst(FirstInst, C, Loc);
2455 if (Check)
2456 Check = ChkBuilder.CreateOr(Check, C);
2457 else
2458 Check = C;
2459 }
2460
2461 // We have to do this trickery because the IRBuilder might fold the check to a
2462 // constant expression in which case there is no Instruction anchored in a
2463 // the block.
2464 LLVMContext &Ctx = Loc->getContext();
2465 Instruction *TheCheck =
2466 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2467 ChkBuilder.Insert(TheCheck, "stride.not.one");
2468 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2469
2470 return std::make_pair(FirstInst, TheCheck);
2471}
2472
2473void InnerLoopVectorizer::createEmptyLoop() {
2474 /*
2475 In this function we generate a new loop. The new loop will contain
2476 the vectorized instructions while the old loop will continue to run the
2477 scalar remainder.
2478
2479 [ ] <-- Back-edge taken count overflow check.
2480 / |
2481 / v
2482 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2483 | / |
2484 | / v
2485 || [ ] <-- vector pre header.
2486 || |
2487 || v
2488 || [ ] \
2489 || [ ]_| <-- vector loop.
2490 || |
2491 | \ v
2492 | >[ ] <--- middle-block.
2493 | / |
2494 | / v
2495 -|- >[ ] <--- new preheader.
2496 | |
2497 | v
2498 | [ ] \
2499 | [ ]_| <-- old scalar loop to handle remainder.
2500 \ |
2501 \ v
2502 >[ ] <-- exit block.
2503 ...
2504 */
2505
2506 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2507 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2508 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2509 assert(BypassBlock && "Invalid loop structure")((BypassBlock && "Invalid loop structure") ? static_cast
<void> (0) : __assert_fail ("BypassBlock && \"Invalid loop structure\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2509, __PRETTY_FUNCTION__));
2510 assert(ExitBlock && "Must have an exit block")((ExitBlock && "Must have an exit block") ? static_cast
<void> (0) : __assert_fail ("ExitBlock && \"Must have an exit block\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2510, __PRETTY_FUNCTION__));
2511
2512 // Some loops have a single integer induction variable, while other loops
2513 // don't. One example is c++ iterators that often have multiple pointer
2514 // induction variables. In the code below we also support a case where we
2515 // don't have a single induction variable.
2516 OldInduction = Legal->getInduction();
2517 Type *IdxTy = Legal->getWidestInductionType();
2518
2519 // Find the loop boundaries.
2520 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2521 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count")((ExitCount != SE->getCouldNotCompute() && "Invalid loop count"
) ? static_cast<void> (0) : __assert_fail ("ExitCount != SE->getCouldNotCompute() && \"Invalid loop count\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2521, __PRETTY_FUNCTION__));
2522
2523 // The exit count might have the type of i64 while the phi is i32. This can
2524 // happen if we have an induction variable that is sign extended before the
2525 // compare. The only way that we get a backedge taken count is that the
2526 // induction variable was signed and as such will not overflow. In such a case
2527 // truncation is legal.
2528 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2529 IdxTy->getPrimitiveSizeInBits())
2530 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2531
2532 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2533 // Get the total trip count from the count by adding 1.
2534 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2535 SE->getConstant(BackedgeTakeCount->getType(), 1));
2536
2537 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2538
2539 // Expand the trip count and place the new instructions in the preheader.
2540 // Notice that the pre-header does not change, only the loop body.
2541 SCEVExpander Exp(*SE, DL, "induction");
2542
2543 // We need to test whether the backedge-taken count is uint##_max. Adding one
2544 // to it will cause overflow and an incorrect loop trip count in the vector
2545 // body. In case of overflow we want to directly jump to the scalar remainder
2546 // loop.
2547 Value *BackedgeCount =
2548 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2549 BypassBlock->getTerminator());
2550 if (BackedgeCount->getType()->isPointerTy())
2551 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2552 "backedge.ptrcnt.to.int",
2553 BypassBlock->getTerminator());
2554 Instruction *CheckBCOverflow =
2555 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2556 Constant::getAllOnesValue(BackedgeCount->getType()),
2557 "backedge.overflow", BypassBlock->getTerminator());
2558
2559 // The loop index does not have to start at Zero. Find the original start
2560 // value from the induction PHI node. If we don't have an induction variable
2561 // then we know that it starts at zero.
2562 Builder.SetInsertPoint(BypassBlock->getTerminator());
2563 Value *StartIdx = ExtendedIdx = OldInduction ?
2564 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2565 IdxTy):
2566 ConstantInt::get(IdxTy, 0);
2567
2568 // We need an instruction to anchor the overflow check on. StartIdx needs to
2569 // be defined before the overflow check branch. Because the scalar preheader
2570 // is going to merge the start index and so the overflow branch block needs to
2571 // contain a definition of the start index.
2572 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2573 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2574 BypassBlock->getTerminator());
2575
2576 // Count holds the overall loop count (N).
2577 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2578 BypassBlock->getTerminator());
2579
2580 LoopBypassBlocks.push_back(BypassBlock);
2581
2582 // Split the single block loop into the two loop structure described above.
2583 BasicBlock *VectorPH =
2584 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2585 BasicBlock *VecBody =
2586 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2587 BasicBlock *MiddleBlock =
2588 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2589 BasicBlock *ScalarPH =
2590 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2591
2592 // Create and register the new vector loop.
2593 Loop* Lp = new Loop();
2594 Loop *ParentLoop = OrigLoop->getParentLoop();
2595
2596 // Insert the new loop into the loop nest and register the new basic blocks
2597 // before calling any utilities such as SCEV that require valid LoopInfo.
2598 if (ParentLoop) {
2599 ParentLoop->addChildLoop(Lp);
2600 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2601 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2602 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2603 } else {
2604 LI->addTopLevelLoop(Lp);
2605 }
2606 Lp->addBasicBlockToLoop(VecBody, *LI);
2607
2608 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2609 // inside the loop.
2610 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2611
2612 // Generate the induction variable.
2613 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2614 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2615 // The loop step is equal to the vectorization factor (num of SIMD elements)
2616 // times the unroll factor (num of SIMD instructions).
2617 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2618
2619 // This is the IR builder that we use to add all of the logic for bypassing
2620 // the new vector loop.
2621 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2622 setDebugLocFromInst(BypassBuilder,
2623 getDebugLocFromInstOrOperands(OldInduction));
2624
2625 // We may need to extend the index in case there is a type mismatch.
2626 // We know that the count starts at zero and does not overflow.
2627 if (Count->getType() != IdxTy) {
2628 // The exit count can be of pointer type. Convert it to the correct
2629 // integer type.
2630 if (ExitCount->getType()->isPointerTy())
2631 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2632 else
2633 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2634 }
2635
2636 // Add the start index to the loop count to get the new end index.
2637 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2638
2639 // Now we need to generate the expression for N - (N % VF), which is
2640 // the part that the vectorized body will execute.
2641 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2642 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2643 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2644 "end.idx.rnd.down");
2645
2646 // Now, compare the new count to zero. If it is zero skip the vector loop and
2647 // jump to the scalar loop.
2648 Value *Cmp =
2649 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2650
2651 BasicBlock *LastBypassBlock = BypassBlock;
2652
2653 // Generate code to check that the loops trip count that we computed by adding
2654 // one to the backedge-taken count will not overflow.
2655 {
2656 auto PastOverflowCheck =
2657 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2658 BasicBlock *CheckBlock =
2659 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2660 if (ParentLoop)
2661 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2662 LoopBypassBlocks.push_back(CheckBlock);
2663 Instruction *OldTerm = LastBypassBlock->getTerminator();
2664 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2665 OldTerm->eraseFromParent();
2666 LastBypassBlock = CheckBlock;
2667 }
2668
2669 // Generate the code to check that the strides we assumed to be one are really
2670 // one. We want the new basic block to start at the first instruction in a
2671 // sequence of instructions that form a check.
2672 Instruction *StrideCheck;
2673 Instruction *FirstCheckInst;
2674 std::tie(FirstCheckInst, StrideCheck) =
2675 addStrideCheck(LastBypassBlock->getTerminator());
2676 if (StrideCheck) {
2677 AddedSafetyChecks = true;
2678 // Create a new block containing the stride check.
2679 BasicBlock *CheckBlock =
2680 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2681 if (ParentLoop)
2682 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2683 LoopBypassBlocks.push_back(CheckBlock);
2684
2685 // Replace the branch into the memory check block with a conditional branch
2686 // for the "few elements case".
2687 Instruction *OldTerm = LastBypassBlock->getTerminator();
2688 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2689 OldTerm->eraseFromParent();
2690
2691 Cmp = StrideCheck;
2692 LastBypassBlock = CheckBlock;
2693 }
2694
2695 // Generate the code that checks in runtime if arrays overlap. We put the
2696 // checks into a separate block to make the more common case of few elements
2697 // faster.
2698 Instruction *MemRuntimeCheck;
2699 std::tie(FirstCheckInst, MemRuntimeCheck) =
2700 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2701 if (MemRuntimeCheck) {
2702 AddedSafetyChecks = true;
2703 // Create a new block containing the memory check.
2704 BasicBlock *CheckBlock =
2705 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2706 if (ParentLoop)
2707 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2708 LoopBypassBlocks.push_back(CheckBlock);
2709
2710 // Replace the branch into the memory check block with a conditional branch
2711 // for the "few elements case".
2712 Instruction *OldTerm = LastBypassBlock->getTerminator();
2713 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2714 OldTerm->eraseFromParent();
2715
2716 Cmp = MemRuntimeCheck;
2717 LastBypassBlock = CheckBlock;
2718 }
2719
2720 LastBypassBlock->getTerminator()->eraseFromParent();
2721 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2722 LastBypassBlock);
2723
2724 // We are going to resume the execution of the scalar loop.
2725 // Go over all of the induction variables that we found and fix the
2726 // PHIs that are left in the scalar version of the loop.
2727 // The starting values of PHI nodes depend on the counter of the last
2728 // iteration in the vectorized loop.
2729 // If we come from a bypass edge then we need to start from the original
2730 // start value.
2731
2732 // This variable saves the new starting index for the scalar loop.
2733 PHINode *ResumeIndex = nullptr;
2734 LoopVectorizationLegality::InductionList::iterator I, E;
2735 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2736 // Set builder to point to last bypass block.
2737 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2738 for (I = List->begin(), E = List->end(); I != E; ++I) {
2739 PHINode *OrigPhi = I->first;
2740 LoopVectorizationLegality::InductionInfo II = I->second;
2741
2742 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2743 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2744 MiddleBlock->getTerminator());
2745 // We might have extended the type of the induction variable but we need a
2746 // truncated version for the scalar loop.
2747 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2748 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2749 MiddleBlock->getTerminator()) : nullptr;
2750
2751 // Create phi nodes to merge from the backedge-taken check block.
2752 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2753 ScalarPH->getTerminator());
2754 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2755
2756 PHINode *BCTruncResumeVal = nullptr;
2757 if (OrigPhi == OldInduction) {
2758 BCTruncResumeVal =
2759 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2760 ScalarPH->getTerminator());
2761 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2762 }
2763
2764 Value *EndValue = nullptr;
2765 switch (II.IK) {
2766 case LoopVectorizationLegality::IK_NoInduction:
2767 llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2767);
2768 case LoopVectorizationLegality::IK_IntInduction: {
2769 // Handle the integer induction counter.
2770 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type")((OrigPhi->getType()->isIntegerTy() && "Invalid type"
) ? static_cast<void> (0) : __assert_fail ("OrigPhi->getType()->isIntegerTy() && \"Invalid type\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2770, __PRETTY_FUNCTION__));
2771
2772 // We have the canonical induction variable.
2773 if (OrigPhi == OldInduction) {
2774 // Create a truncated version of the resume value for the scalar loop,
2775 // we might have promoted the type to a larger width.
2776 EndValue =
2777 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2778 // The new PHI merges the original incoming value, in case of a bypass,
2779 // or the value at the end of the vectorized loop.
2780 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2781 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2782 TruncResumeVal->addIncoming(EndValue, VecBody);
2783
2784 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2785
2786 // We know what the end value is.
2787 EndValue = IdxEndRoundDown;
2788 // We also know which PHI node holds it.
2789 ResumeIndex = ResumeVal;
2790 break;
2791 }
2792
2793 // Not the canonical induction variable - add the vector loop count to the
2794 // start value.
2795 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2796 II.StartValue->getType(),
2797 "cast.crd");
2798 EndValue = II.transform(BypassBuilder, CRD);
2799 EndValue->setName("ind.end");
2800 break;
2801 }
2802 case LoopVectorizationLegality::IK_PtrInduction: {
2803 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2804 II.StepValue->getType(),
2805 "cast.crd");
2806 EndValue = II.transform(BypassBuilder, CRD);
2807 EndValue->setName("ptr.ind.end");
2808 break;
2809 }
2810 }// end of case
2811
2812 // The new PHI merges the original incoming value, in case of a bypass,
2813 // or the value at the end of the vectorized loop.
2814 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2815 if (OrigPhi == OldInduction)
2816 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2817 else
2818 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2819 }
2820 ResumeVal->addIncoming(EndValue, VecBody);
2821
2822 // Fix the scalar body counter (PHI node).
2823 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2824
2825 // The old induction's phi node in the scalar body needs the truncated
2826 // value.
2827 if (OrigPhi == OldInduction) {
2828 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2829 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2830 } else {
2831 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2832 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2833 }
2834 }
2835
2836 // If we are generating a new induction variable then we also need to
2837 // generate the code that calculates the exit value. This value is not
2838 // simply the end of the counter because we may skip the vectorized body
2839 // in case of a runtime check.
2840 if (!OldInduction){
2841 assert(!ResumeIndex && "Unexpected resume value found")((!ResumeIndex && "Unexpected resume value found") ? static_cast
<void> (0) : __assert_fail ("!ResumeIndex && \"Unexpected resume value found\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2841, __PRETTY_FUNCTION__));
2842 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2843 MiddleBlock->getTerminator());
2844 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2845 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2846 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2847 }
2848
2849 // Make sure that we found the index where scalar loop needs to continue.
2850 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&((ResumeIndex && ResumeIndex->getType()->isIntegerTy
() && "Invalid resume Index") ? static_cast<void>
(0) : __assert_fail ("ResumeIndex && ResumeIndex->getType()->isIntegerTy() && \"Invalid resume Index\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2851, __PRETTY_FUNCTION__))
2851 "Invalid resume Index")((ResumeIndex && ResumeIndex->getType()->isIntegerTy
() && "Invalid resume Index") ? static_cast<void>
(0) : __assert_fail ("ResumeIndex && ResumeIndex->getType()->isIntegerTy() && \"Invalid resume Index\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2851, __PRETTY_FUNCTION__));
2852
2853 // Add a check in the middle block to see if we have completed
2854 // all of the iterations in the first vector loop.
2855 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2856 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2857 ResumeIndex, "cmp.n",
2858 MiddleBlock->getTerminator());
2859
2860 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2861 // Remove the old terminator.
2862 MiddleBlock->getTerminator()->eraseFromParent();
2863
2864 // Create i+1 and fill the PHINode.
2865 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2866 Induction->addIncoming(StartIdx, VectorPH);
2867 Induction->addIncoming(NextIdx, VecBody);
2868 // Create the compare.
2869 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2870 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2871
2872 // Now we have two terminators. Remove the old one from the block.
2873 VecBody->getTerminator()->eraseFromParent();
2874
2875 // Get ready to start creating new instructions into the vectorized body.
2876 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2877
2878 // Save the state.
2879 LoopVectorPreHeader = VectorPH;
2880 LoopScalarPreHeader = ScalarPH;
2881 LoopMiddleBlock = MiddleBlock;
2882 LoopExitBlock = ExitBlock;
2883 LoopVectorBody.push_back(VecBody);
2884 LoopScalarBody = OldBasicBlock;
2885
2886 LoopVectorizeHints Hints(Lp, true);
2887 Hints.setAlreadyVectorized();
2888}
2889
2890namespace {
2891struct CSEDenseMapInfo {
2892 static bool canHandle(Instruction *I) {
2893 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2894 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2895 }
2896 static inline Instruction *getEmptyKey() {
2897 return DenseMapInfo<Instruction *>::getEmptyKey();
2898 }
2899 static inline Instruction *getTombstoneKey() {
2900 return DenseMapInfo<Instruction *>::getTombstoneKey();
2901 }
2902 static unsigned getHashValue(Instruction *I) {
2903 assert(canHandle(I) && "Unknown instruction!")((canHandle(I) && "Unknown instruction!") ? static_cast
<void> (0) : __assert_fail ("canHandle(I) && \"Unknown instruction!\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2903, __PRETTY_FUNCTION__));
2904 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2905 I->value_op_end()));
2906 }
2907 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2908 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2909 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2910 return LHS == RHS;
2911 return LHS->isIdenticalTo(RHS);
2912 }
2913};
2914}
2915
2916/// \brief Check whether this block is a predicated block.
2917/// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2918/// = ...; " blocks. We start with one vectorized basic block. For every
2919/// conditional block we split this vectorized block. Therefore, every second
2920/// block will be a predicated one.
2921static bool isPredicatedBlock(unsigned BlockNum) {
2922 return BlockNum % 2;
2923}
2924
2925///\brief Perform cse of induction variable instructions.
2926static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2927 // Perform simple cse.
2928 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2929 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2930 BasicBlock *BB = BBs[i];
2931 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2932 Instruction *In = I++;
2933
2934 if (!CSEDenseMapInfo::canHandle(In))
2935 continue;
2936
2937 // Check if we can replace this instruction with any of the
2938 // visited instructions.
2939 if (Instruction *V = CSEMap.lookup(In)) {
2940 In->replaceAllUsesWith(V);
2941 In->eraseFromParent();
2942 continue;
2943 }
2944 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2945 // ...;" blocks for predicated stores. Every second block is a predicated
2946 // block.
2947 if (isPredicatedBlock(i))
2948 continue;
2949
2950 CSEMap[In] = In;
2951 }
2952 }
2953}
2954
2955/// \brief Adds a 'fast' flag to floating point operations.
2956static Value *addFastMathFlag(Value *V) {
2957 if (isa<FPMathOperator>(V)){
2958 FastMathFlags Flags;
2959 Flags.setUnsafeAlgebra();
2960 cast<Instruction>(V)->setFastMathFlags(Flags);
2961 }
2962 return V;
2963}
2964
2965/// Estimate the overhead of scalarizing a value. Insert and Extract are set if
2966/// the result needs to be inserted and/or extracted from vectors.
2967static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
2968 const TargetTransformInfo &TTI) {
2969 if (Ty->isVoidTy())
2970 return 0;
2971
2972 assert(Ty->isVectorTy() && "Can only scalarize vectors")((Ty->isVectorTy() && "Can only scalarize vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only scalarize vectors\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2972, __PRETTY_FUNCTION__));
2973 unsigned Cost = 0;
2974
2975 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
2976 if (Insert)
2977 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
2978 if (Extract)
2979 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
2980 }
2981
2982 return Cost;
2983}
2984
2985// Estimate cost of a call instruction CI if it were vectorized with factor VF.
2986// Return the cost of the instruction, including scalarization overhead if it's
2987// needed. The flag NeedToScalarize shows if the call needs to be scalarized -
2988// i.e. either vector version isn't available, or is too expensive.
2989static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
2990 const TargetTransformInfo &TTI,
2991 const TargetLibraryInfo *TLI,
2992 bool &NeedToScalarize) {
2993 Function *F = CI->getCalledFunction();
2994 StringRef FnName = CI->getCalledFunction()->getName();
2995 Type *ScalarRetTy = CI->getType();
2996 SmallVector<Type *, 4> Tys, ScalarTys;
2997 for (auto &ArgOp : CI->arg_operands())
2998 ScalarTys.push_back(ArgOp->getType());
2999
3000 // Estimate cost of scalarized vector call. The source operands are assumed
3001 // to be vectors, so we need to extract individual elements from there,
3002 // execute VF scalar calls, and then gather the result into the vector return
3003 // value.
3004 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3005 if (VF == 1)
3006 return ScalarCallCost;
3007
3008 // Compute corresponding vector type for return value and arguments.
3009 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3010 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3011 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3012
3013 // Compute costs of unpacking argument values for the scalar calls and
3014 // packing the return values to a vector.
3015 unsigned ScalarizationCost =
3016 getScalarizationOverhead(RetTy, true, false, TTI);
3017 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3018 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3019
3020 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3021
3022 // If we can't emit a vector call for this function, then the currently found
3023 // cost is the cost we need to return.
3024 NeedToScalarize = true;
3025 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3026 return Cost;
3027
3028 // If the corresponding vector cost is cheaper, return its cost.
3029 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3030 if (VectorCallCost < Cost) {
3031 NeedToScalarize = false;
3032 return VectorCallCost;
3033 }
3034 return Cost;
3035}
3036
3037// Estimate cost of an intrinsic call instruction CI if it were vectorized with
3038// factor VF. Return the cost of the instruction, including scalarization
3039// overhead if it's needed.
3040static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3041 const TargetTransformInfo &TTI,
3042 const TargetLibraryInfo *TLI) {
3043 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3044 assert(ID && "Expected intrinsic call!")((ID && "Expected intrinsic call!") ? static_cast<
void> (0) : __assert_fail ("ID && \"Expected intrinsic call!\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3044, __PRETTY_FUNCTION__));
3045
3046 Type *RetTy = ToVectorTy(CI->getType(), VF);
3047 SmallVector<Type *, 4> Tys;
3048 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3049 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3050
3051 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3052}
3053
3054void InnerLoopVectorizer::vectorizeLoop() {
3055 //===------------------------------------------------===//
3056 //
3057 // Notice: any optimization or new instruction that go
3058 // into the code below should be also be implemented in
3059 // the cost-model.
3060 //
3061 //===------------------------------------------------===//
3062 Constant *Zero = Builder.getInt32(0);
3063
3064 // In order to support reduction variables we need to be able to vectorize
3065 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3066 // stages. First, we create a new vector PHI node with no incoming edges.
3067 // We use this value when we vectorize all of the instructions that use the
3068 // PHI. Next, after all of the instructions in the block are complete we
3069 // add the new incoming edges to the PHI. At this point all of the
3070 // instructions in the basic block are vectorized, so we can use them to
3071 // construct the PHI.
3072 PhiVector RdxPHIsToFix;
3073
3074 // Scan the loop in a topological order to ensure that defs are vectorized
3075 // before users.
3076 LoopBlocksDFS DFS(OrigLoop);
3077 DFS.perform(LI);
3078
3079 // Vectorize all of the blocks in the original loop.
3080 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3081 be = DFS.endRPO(); bb != be; ++bb)
3082 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3083
3084 // At this point every instruction in the original loop is widened to
3085 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3086 // that we vectorized. The PHI nodes are currently empty because we did
3087 // not want to introduce cycles. Notice that the remaining PHI nodes
3088 // that we need to fix are reduction variables.
3089
3090 // Create the 'reduced' values for each of the induction vars.
3091 // The reduced values are the vector values that we scalarize and combine
3092 // after the loop is finished.
3093 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3094 it != e; ++it) {
3095 PHINode *RdxPhi = *it;
3096 assert(RdxPhi && "Unable to recover vectorized PHI")((RdxPhi && "Unable to recover vectorized PHI") ? static_cast
<void> (0) : __assert_fail ("RdxPhi && \"Unable to recover vectorized PHI\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3096, __PRETTY_FUNCTION__));
3097
3098 // Find the reduction variable descriptor.
3099 assert(Legal->getReductionVars()->count(RdxPhi) &&((Legal->getReductionVars()->count(RdxPhi) && "Unable to find the reduction variable"
) ? static_cast<void> (0) : __assert_fail ("Legal->getReductionVars()->count(RdxPhi) && \"Unable to find the reduction variable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3100, __PRETTY_FUNCTION__))
3100 "Unable to find the reduction variable")((Legal->getReductionVars()->count(RdxPhi) && "Unable to find the reduction variable"
) ? static_cast<void> (0) : __assert_fail ("Legal->getReductionVars()->count(RdxPhi) && \"Unable to find the reduction variable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3100, __PRETTY_FUNCTION__));
3101 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3102
3103 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3104 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3105 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3106 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3107 RdxDesc.getMinMaxRecurrenceKind();
3108 setDebugLocFromInst(Builder, ReductionStartValue);
3109
3110 // We need to generate a reduction vector from the incoming scalar.
3111 // To do so, we need to generate the 'identity' vector and override
3112 // one of the elements with the incoming scalar reduction. We need
3113 // to do it in the vector-loop preheader.
3114 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3115
3116 // This is the vector-clone of the value that leaves the loop.
3117 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3118 Type *VecTy = VectorExit[0]->getType();
3119
3120 // Find the reduction identity variable. Zero for addition, or, xor,
3121 // one for multiplication, -1 for And.
3122 Value *Identity;
3123 Value *VectorStart;
3124 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3125 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3126 // MinMax reduction have the start value as their identify.
3127 if (VF == 1) {
3128 VectorStart = Identity = ReductionStartValue;
3129 } else {
3130 VectorStart = Identity =
3131 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3132 }
3133 } else {
3134 // Handle other reduction kinds:
3135 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3136 RK, VecTy->getScalarType());
3137 if (VF == 1) {
3138 Identity = Iden;
3139 // This vector is the Identity vector where the first element is the
3140 // incoming scalar reduction.
3141 VectorStart = ReductionStartValue;
3142 } else {
3143 Identity = ConstantVector::getSplat(VF, Iden);
3144
3145 // This vector is the Identity vector where the first element is the
3146 // incoming scalar reduction.
3147 VectorStart =
3148 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3149 }
3150 }
3151
3152 // Fix the vector-loop phi.
3153
3154 // Reductions do not have to start at zero. They can start with
3155 // any loop invariant values.
3156 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3157 BasicBlock *Latch = OrigLoop->getLoopLatch();
3158 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3159 VectorParts &Val = getVectorValue(LoopVal);
3160 for (unsigned part = 0; part < UF; ++part) {
3161 // Make sure to add the reduction stat value only to the
3162 // first unroll part.
3163 Value *StartVal = (part == 0) ? VectorStart : Identity;
3164 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3165 LoopVectorPreHeader);
3166 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3167 LoopVectorBody.back());
3168 }
3169
3170 // Before each round, move the insertion point right between
3171 // the PHIs and the values we are going to write.
3172 // This allows us to write both PHINodes and the extractelement
3173 // instructions.
3174 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3175
3176 VectorParts RdxParts;
3177 setDebugLocFromInst(Builder, LoopExitInst);
3178 for (unsigned part = 0; part < UF; ++part) {
3179 // This PHINode contains the vectorized reduction variable, or
3180 // the initial value vector, if we bypass the vector loop.
3181 VectorParts &RdxExitVal = getVectorValue(LoopExitInst);
3182 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3183 Value *StartVal = (part == 0) ? VectorStart : Identity;
3184 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3185 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3186 NewPhi->addIncoming(RdxExitVal[part],
3187 LoopVectorBody.back());
3188 RdxParts.push_back(NewPhi);
3189 }
3190
3191 // Reduce all of the unrolled parts into a single vector.
3192 Value *ReducedPartRdx = RdxParts[0];
3193 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3194 setDebugLocFromInst(Builder, ReducedPartRdx);
3195 for (unsigned part = 1; part < UF; ++part) {
3196 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3197 // Floating point operations had to be 'fast' to enable the reduction.
3198 ReducedPartRdx = addFastMathFlag(
3199 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3200 ReducedPartRdx, "bin.rdx"));
3201 else
3202 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3203 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3204 }
3205
3206 if (VF > 1) {
3207 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3208 // and vector ops, reducing the set of values being computed by half each
3209 // round.
3210 assert(isPowerOf2_32(VF) &&((isPowerOf2_32(VF) && "Reduction emission only supported for pow2 vectors!"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF) && \"Reduction emission only supported for pow2 vectors!\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3211, __PRETTY_FUNCTION__))
3211 "Reduction emission only supported for pow2 vectors!")((isPowerOf2_32(VF) && "Reduction emission only supported for pow2 vectors!"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(VF) && \"Reduction emission only supported for pow2 vectors!\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3211, __PRETTY_FUNCTION__));
3212 Value *TmpVec = ReducedPartRdx;
3213 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3214 for (unsigned i = VF; i != 1; i >>= 1) {
3215 // Move the upper half of the vector to the lower half.
3216 for (unsigned j = 0; j != i/2; ++j)
3217 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3218
3219 // Fill the rest of the mask with undef.
3220 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3221 UndefValue::get(Builder.getInt32Ty()));
3222
3223 Value *Shuf =
3224 Builder.CreateShuffleVector(TmpVec,
3225 UndefValue::get(TmpVec->getType()),
3226 ConstantVector::get(ShuffleMask),
3227 "rdx.shuf");
3228
3229 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3230 // Floating point operations had to be 'fast' to enable the reduction.
3231 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3232 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3233 else
3234 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3235 TmpVec, Shuf);
3236 }
3237
3238 // The result is in the first element of the vector.
3239 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3240 Builder.getInt32(0));
3241 }
3242
3243 // Create a phi node that merges control-flow from the backedge-taken check
3244 // block and the middle block.
3245 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3246 LoopScalarPreHeader->getTerminator());
3247 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3248 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3249
3250 // Now, we need to fix the users of the reduction variable
3251 // inside and outside of the scalar remainder loop.
3252 // We know that the loop is in LCSSA form. We need to update the
3253 // PHI nodes in the exit blocks.
3254 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3255 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3256 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3257 if (!LCSSAPhi) break;
3258
3259 // All PHINodes need to have a single entry edge, or two if
3260 // we already fixed them.
3261 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI")((LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"
) ? static_cast<void> (0) : __assert_fail ("LCSSAPhi->getNumIncomingValues() < 3 && \"Invalid LCSSA PHI\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3261, __PRETTY_FUNCTION__));
3262
3263 // We found our reduction value exit-PHI. Update it with the
3264 // incoming bypass edge.
3265 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3266 // Add an edge coming from the bypass.
3267 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3268 break;
3269 }
3270 }// end of the LCSSA phi scan.
3271
3272 // Fix the scalar loop reduction variable with the incoming reduction sum
3273 // from the vector body and from the backedge value.
3274 int IncomingEdgeBlockIdx =
3275 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3276 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index")((IncomingEdgeBlockIdx >= 0 && "Invalid block index"
) ? static_cast<void> (0) : __assert_fail ("IncomingEdgeBlockIdx >= 0 && \"Invalid block index\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3276, __PRETTY_FUNCTION__));
3277 // Pick the other block.
3278 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3279 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3280 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3281 }// end of for each redux variable.
3282
3283 fixLCSSAPHIs();
3284
3285 // Remove redundant induction instructions.
3286 cse(LoopVectorBody);
3287}
3288
3289void InnerLoopVectorizer::fixLCSSAPHIs() {
3290 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3291 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3292 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3293 if (!LCSSAPhi) break;
3294 if (LCSSAPhi->getNumIncomingValues() == 1)
3295 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3296 LoopMiddleBlock);
3297 }
3298}
3299
3300InnerLoopVectorizer::VectorParts
3301InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3302 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&((std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(
Dst) && "Invalid edge") ? static_cast<void> (0)
: __assert_fail ("std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && \"Invalid edge\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3303, __PRETTY_FUNCTION__))
3303 "Invalid edge")((std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(
Dst) && "Invalid edge") ? static_cast<void> (0)
: __assert_fail ("std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && \"Invalid edge\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3303, __PRETTY_FUNCTION__));
3304
3305 // Look for cached value.
3306 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3307 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3308 if (ECEntryIt != MaskCache.end())
3309 return ECEntryIt->second;
3310
3311 VectorParts SrcMask = createBlockInMask(Src);
3312
3313 // The terminator has to be a branch inst!
3314 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3315 assert(BI && "Unexpected terminator found")((BI && "Unexpected terminator found") ? static_cast<
void> (0) : __assert_fail ("BI && \"Unexpected terminator found\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3315, __PRETTY_FUNCTION__));
3316
3317 if (BI->isConditional()) {
3318 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3319
3320 if (BI->getSuccessor(0) != Dst)
3321 for (unsigned part = 0; part < UF; ++part)
3322 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3323
3324 for (unsigned part = 0; part < UF; ++part)
3325 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3326
3327 MaskCache[Edge] = EdgeMask;
3328 return EdgeMask;
3329 }
3330
3331 MaskCache[Edge] = SrcMask;
3332 return SrcMask;
3333}
3334
3335InnerLoopVectorizer::VectorParts
3336InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3337 assert(OrigLoop->contains(BB) && "Block is not a part of a loop")((OrigLoop->contains(BB) && "Block is not a part of a loop"
) ? static_cast<void> (0) : __assert_fail ("OrigLoop->contains(BB) && \"Block is not a part of a loop\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3337, __PRETTY_FUNCTION__));
3338
3339 // Loop incoming mask is all-one.
3340 if (OrigLoop->getHeader() == BB) {
3341 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3342 return getVectorValue(C);
3343 }
3344
3345 // This is the block mask. We OR all incoming edges, and with zero.
3346 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3347 VectorParts BlockMask = getVectorValue(Zero);
3348
3349 // For each pred:
3350 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3351 VectorParts EM = createEdgeMask(*it, BB);
3352 for (unsigned part = 0; part < UF; ++part)
3353 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3354 }
3355
3356 return BlockMask;
3357}
3358
3359void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3360 InnerLoopVectorizer::VectorParts &Entry,
3361 unsigned UF, unsigned VF, PhiVector *PV) {
3362 PHINode* P = cast<PHINode>(PN);
3363 // Handle reduction variables:
3364 if (Legal->getReductionVars()->count(P)) {
3365 for (unsigned part = 0; part < UF; ++part) {
3366 // This is phase one of vectorizing PHIs.
3367 Type *VecTy = (VF == 1) ? PN->getType() :
3368 VectorType::get(PN->getType(), VF);
3369 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3370 LoopVectorBody.back()-> getFirstInsertionPt());
3371 }
3372 PV->push_back(P);
3373 return;
3374 }
3375
3376 setDebugLocFromInst(Builder, P);
3377 // Check for PHI nodes that are lowered to vector selects.
3378 if (P->getParent() != OrigLoop->getHeader()) {
3379 // We know that all PHIs in non-header blocks are converted into
3380 // selects, so we don't have to worry about the insertion order and we
3381 // can just use the builder.
3382 // At this point we generate the predication tree. There may be
3383 // duplications since this is a simple recursive scan, but future
3384 // optimizations will clean it up.
3385
3386 unsigned NumIncoming = P->getNumIncomingValues();
3387
3388 // Generate a sequence of selects of the form:
3389 // SELECT(Mask3, In3,
3390 // SELECT(Mask2, In2,
3391 // ( ...)))
3392 for (unsigned In = 0; In < NumIncoming; In++) {
3393 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3394 P->getParent());
3395 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3396
3397 for (unsigned part = 0; part < UF; ++part) {
3398 // We might have single edge PHIs (blocks) - use an identity
3399 // 'select' for the first PHI operand.
3400 if (In == 0)
3401 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3402 In0[part]);
3403 else
3404 // Select between the current value and the previous incoming edge
3405 // based on the incoming mask.
3406 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3407 Entry[part], "predphi");
3408 }
3409 }
3410 return;
3411 }
3412
3413 // This PHINode must be an induction variable.
3414 // Make sure that we know about it.
3415 assert(Legal->getInductionVars()->count(P) &&((Legal->getInductionVars()->count(P) && "Not an induction variable"
) ? static_cast<void> (0) : __assert_fail ("Legal->getInductionVars()->count(P) && \"Not an induction variable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3416, __PRETTY_FUNCTION__))
3416 "Not an induction variable")((Legal->getInductionVars()->count(P) && "Not an induction variable"
) ? static_cast<void> (0) : __assert_fail ("Legal->getInductionVars()->count(P) && \"Not an induction variable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3416, __PRETTY_FUNCTION__));
3417
3418 LoopVectorizationLegality::InductionInfo II =
3419 Legal->getInductionVars()->lookup(P);
3420
3421 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3422 // which can be found from the original scalar operations.
3423 switch (II.IK) {
3424 case LoopVectorizationLegality::IK_NoInduction:
3425 llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3425);
3426 case LoopVectorizationLegality::IK_IntInduction: {
3427 assert(P->getType() == II.StartValue->getType() && "Types must match")((P->getType() == II.StartValue->getType() && "Types must match"
) ? static_cast<void> (0) : __assert_fail ("P->getType() == II.StartValue->getType() && \"Types must match\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3427, __PRETTY_FUNCTION__));
3428 Type *PhiTy = P->getType();
3429 Value *Broadcasted;
3430 if (P == OldInduction) {
3431 // Handle the canonical induction variable. We might have had to
3432 // extend the type.
3433 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3434 } else {
3435 // Handle other induction variables that are now based on the
3436 // canonical one.
3437 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3438 "normalized.idx");
3439 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3440 Broadcasted = II.transform(Builder, NormalizedIdx);
3441 Broadcasted->setName("offset.idx");
3442 }
3443 Broadcasted = getBroadcastInstrs(Broadcasted);
3444 // After broadcasting the induction variable we need to make the vector
3445 // consecutive by adding 0, 1, 2, etc.
3446 for (unsigned part = 0; part < UF; ++part)
3447 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3448 return;
3449 }
3450 case LoopVectorizationLegality::IK_PtrInduction:
3451 // Handle the pointer induction variable case.
3452 assert(P->getType()->isPointerTy() && "Unexpected type.")((P->getType()->isPointerTy() && "Unexpected type."
) ? static_cast<void> (0) : __assert_fail ("P->getType()->isPointerTy() && \"Unexpected type.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3452, __PRETTY_FUNCTION__));
3453 // This is the normalized GEP that starts counting at zero.
3454 Value *NormalizedIdx =
3455 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3456 NormalizedIdx =
3457 Builder.CreateSExtOrTrunc(NormalizedIdx, II.StepValue->getType());
3458 // This is the vector of results. Notice that we don't generate
3459 // vector geps because scalar geps result in better code.
3460 for (unsigned part = 0; part < UF; ++part) {
3461 if (VF == 1) {
3462 int EltIndex = part;
3463 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3464 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3465 Value *SclrGep = II.transform(Builder, GlobalIdx);
3466 SclrGep->setName("next.gep");
3467 Entry[part] = SclrGep;
3468 continue;
3469 }
3470
3471 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3472 for (unsigned int i = 0; i < VF; ++i) {
3473 int EltIndex = i + part * VF;
3474 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3475 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3476 Value *SclrGep = II.transform(Builder, GlobalIdx);
3477 SclrGep->setName("next.gep");
3478 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3479 Builder.getInt32(i),
3480 "insert.gep");
3481 }
3482 Entry[part] = VecVal;
3483 }
3484 return;
3485 }
3486}
3487
3488void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3489 // For each instruction in the old loop.
3490 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3491 VectorParts &Entry = WidenMap.get(it);
3492 switch (it->getOpcode()) {
3493 case Instruction::Br:
3494 // Nothing to do for PHIs and BR, since we already took care of the
3495 // loop control flow instructions.
3496 continue;
3497 case Instruction::PHI: {
3498 // Vectorize PHINodes.
3499 widenPHIInstruction(it, Entry, UF, VF, PV);
3500 continue;
3501 }// End of PHI.
3502
3503 case Instruction::Add:
3504 case Instruction::FAdd:
3505 case Instruction::Sub:
3506 case Instruction::FSub:
3507 case Instruction::Mul:
3508 case Instruction::FMul:
3509 case Instruction::UDiv:
3510 case Instruction::SDiv:
3511 case Instruction::FDiv:
3512 case Instruction::URem:
3513 case Instruction::SRem:
3514 case Instruction::FRem:
3515 case Instruction::Shl:
3516 case Instruction::LShr:
3517 case Instruction::AShr:
3518 case Instruction::And:
3519 case Instruction::Or:
3520 case Instruction::Xor: {
3521 // Just widen binops.
3522 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3523 setDebugLocFromInst(Builder, BinOp);
3524 VectorParts &A = getVectorValue(it->getOperand(0));
3525 VectorParts &B = getVectorValue(it->getOperand(1));
3526
3527 // Use this vector value for all users of the original instruction.
3528 for (unsigned Part = 0; Part < UF; ++Part) {
3529 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3530
3531 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3532 VecOp->copyIRFlags(BinOp);
3533
3534 Entry[Part] = V;
3535 }
3536
3537 propagateMetadata(Entry, it);
3538 break;
3539 }
3540 case Instruction::Select: {
3541 // Widen selects.
3542 // If the selector is loop invariant we can create a select
3543 // instruction with a scalar condition. Otherwise, use vector-select.
3544 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3545 OrigLoop);
3546 setDebugLocFromInst(Builder, it);
3547
3548 // The condition can be loop invariant but still defined inside the
3549 // loop. This means that we can't just use the original 'cond' value.
3550 // We have to take the 'vectorized' value and pick the first lane.
3551 // Instcombine will make this a no-op.
3552 VectorParts &Cond = getVectorValue(it->getOperand(0));
3553 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3554 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3555
3556 Value *ScalarCond = (VF == 1) ? Cond[0] :
3557 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3558
3559 for (unsigned Part = 0; Part < UF; ++Part) {
3560 Entry[Part] = Builder.CreateSelect(
3561 InvariantCond ? ScalarCond : Cond[Part],
3562 Op0[Part],
3563 Op1[Part]);
3564 }
3565
3566 propagateMetadata(Entry, it);
3567 break;
3568 }
3569
3570 case Instruction::ICmp:
3571 case Instruction::FCmp: {
3572 // Widen compares. Generate vector compares.
3573 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3574 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3575 setDebugLocFromInst(Builder, it);
3576 VectorParts &A = getVectorValue(it->getOperand(0));
3577 VectorParts &B = getVectorValue(it->getOperand(1));
3578 for (unsigned Part = 0; Part < UF; ++Part) {
3579 Value *C = nullptr;
3580 if (FCmp)
3581 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3582 else
3583 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3584 Entry[Part] = C;
3585 }
3586
3587 propagateMetadata(Entry, it);
3588 break;
3589 }
3590
3591 case Instruction::Store:
3592 case Instruction::Load:
3593 vectorizeMemoryInstruction(it);
3594 break;
3595 case Instruction::ZExt:
3596 case Instruction::SExt:
3597 case Instruction::FPToUI:
3598 case Instruction::FPToSI:
3599 case Instruction::FPExt:
3600 case Instruction::PtrToInt:
3601 case Instruction::IntToPtr:
3602 case Instruction::SIToFP:
3603 case Instruction::UIToFP:
3604 case Instruction::Trunc:
3605 case Instruction::FPTrunc:
3606 case Instruction::BitCast: {
3607 CastInst *CI = dyn_cast<CastInst>(it);
3608 setDebugLocFromInst(Builder, it);
3609 /// Optimize the special case where the source is the induction
3610 /// variable. Notice that we can only optimize the 'trunc' case
3611 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3612 /// c. other casts depend on pointer size.
3613 if (CI->getOperand(0) == OldInduction &&
3614 it->getOpcode() == Instruction::Trunc) {
3615 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3616 CI->getType());
3617 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3618 LoopVectorizationLegality::InductionInfo II =
3619 Legal->getInductionVars()->lookup(OldInduction);
3620 Constant *Step =
3621 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3622 for (unsigned Part = 0; Part < UF; ++Part)
3623 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3624 propagateMetadata(Entry, it);
3625 break;
3626 }
3627 /// Vectorize casts.
3628 Type *DestTy = (VF == 1) ? CI->getType() :
3629 VectorType::get(CI->getType(), VF);
3630
3631 VectorParts &A = getVectorValue(it->getOperand(0));
3632 for (unsigned Part = 0; Part < UF; ++Part)
3633 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3634 propagateMetadata(Entry, it);
3635 break;
3636 }
3637
3638 case Instruction::Call: {
3639 // Ignore dbg intrinsics.
3640 if (isa<DbgInfoIntrinsic>(it))
3641 break;
3642 setDebugLocFromInst(Builder, it);
3643
3644 Module *M = BB->getParent()->getParent();
3645 CallInst *CI = cast<CallInst>(it);
3646
3647 StringRef FnName = CI->getCalledFunction()->getName();
3648 Function *F = CI->getCalledFunction();
3649 Type *RetTy = ToVectorTy(CI->getType(), VF);
3650 SmallVector<Type *, 4> Tys;
3651 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3652 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3653
3654 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3655 if (ID &&
3656 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3657 ID == Intrinsic::lifetime_start)) {
3658 scalarizeInstruction(it);
3659 break;
3660 }
3661 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3662 // version of the instruction.
3663 // Is it beneficial to perform intrinsic call compared to lib call?
3664 bool NeedToScalarize;
3665 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3666 bool UseVectorIntrinsic =
3667 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3668 if (!UseVectorIntrinsic && NeedToScalarize) {
3669 scalarizeInstruction(it);
3670 break;
3671 }
3672
3673 for (unsigned Part = 0; Part < UF; ++Part) {
3674 SmallVector<Value *, 4> Args;
3675 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3676 Value *Arg = CI->getArgOperand(i);
3677 // Some intrinsics have a scalar argument - don't replace it with a
3678 // vector.
3679 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3680 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3681 Arg = VectorArg[Part];
3682 }
3683 Args.push_back(Arg);
3684 }
3685
3686 Function *VectorF;
3687 if (UseVectorIntrinsic) {
3688 // Use vector version of the intrinsic.
3689 Type *TysForDecl[] = {CI->getType()};
3690 if (VF > 1)
3691 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3692 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3693 } else {
3694 // Use vector version of the library call.
3695 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3696 assert(!VFnName.empty() && "Vector function name is empty.")((!VFnName.empty() && "Vector function name is empty."
) ? static_cast<void> (0) : __assert_fail ("!VFnName.empty() && \"Vector function name is empty.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3696, __PRETTY_FUNCTION__));
3697 VectorF = M->getFunction(VFnName);
3698 if (!VectorF) {
3699 // Generate a declaration
3700 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3701 VectorF =
3702 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3703 VectorF->copyAttributesFrom(F);
3704 }
3705 }
3706 assert(VectorF && "Can't create vector function.")((VectorF && "Can't create vector function.") ? static_cast
<void> (0) : __assert_fail ("VectorF && \"Can't create vector function.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3706, __PRETTY_FUNCTION__));
3707 Entry[Part] = Builder.CreateCall(VectorF, Args);
3708 }
3709
3710 propagateMetadata(Entry, it);
3711 break;
3712 }
3713
3714 default:
3715 // All other instructions are unsupported. Scalarize them.
3716 scalarizeInstruction(it);
3717 break;
3718 }// end of switch.
3719 }// end of for_each instr.
3720}
3721
3722void InnerLoopVectorizer::updateAnalysis() {
3723 // Forget the original basic block.
3724 SE->forgetLoop(OrigLoop);
3725
3726 // Update the dominator tree information.
3727 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&((DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock
) && "Entry does not dominate exit.") ? static_cast<
void> (0) : __assert_fail ("DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && \"Entry does not dominate exit.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3728, __PRETTY_FUNCTION__))
3728 "Entry does not dominate exit.")((DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock
) && "Entry does not dominate exit.") ? static_cast<
void> (0) : __assert_fail ("DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && \"Entry does not dominate exit.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3728, __PRETTY_FUNCTION__));
3729
3730 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3731 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3732 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3733
3734 // Due to if predication of stores we might create a sequence of "if(pred)
3735 // a[i] = ...; " blocks.
3736 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3737 if (i == 0)
3738 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3739 else if (isPredicatedBlock(i)) {
3740 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3741 } else {
3742 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3743 }
3744 }
3745
3746 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3747 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3748 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3749 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3750
3751 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (0);
3752}
3753
3754/// \brief Check whether it is safe to if-convert this phi node.
3755///
3756/// Phi nodes with constant expressions that can trap are not safe to if
3757/// convert.
3758static bool canIfConvertPHINodes(BasicBlock *BB) {
3759 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3760 PHINode *Phi = dyn_cast<PHINode>(I);
3761 if (!Phi)
3762 return true;
3763 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3764 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3765 if (C->canTrap())
3766 return false;
3767 }
3768 return true;
3769}
3770
3771bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3772 if (!EnableIfConversion) {
3773 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3774 return false;
3775 }
3776
3777 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable")((TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"
) ? static_cast<void> (0) : __assert_fail ("TheLoop->getNumBlocks() > 1 && \"Single block loops are vectorizable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3777, __PRETTY_FUNCTION__));
3778
3779 // A list of pointers that we can safely read and write to.
3780 SmallPtrSet<Value *, 8> SafePointes;
3781
3782 // Collect safe addresses.
3783 for (Loop::block_iterator BI = TheLoop->block_begin(),
3784 BE = TheLoop->block_end(); BI != BE; ++BI) {
3785 BasicBlock *BB = *BI;
3786
3787 if (blockNeedsPredication(BB))
3788 continue;
3789
3790 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3791 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3792 SafePointes.insert(LI->getPointerOperand());
3793 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3794 SafePointes.insert(SI->getPointerOperand());
3795 }
3796 }
3797
3798 // Collect the blocks that need predication.
3799 BasicBlock *Header = TheLoop->getHeader();
3800 for (Loop::block_iterator BI = TheLoop->block_begin(),
3801 BE = TheLoop->block_end(); BI != BE; ++BI) {
3802 BasicBlock *BB = *BI;
3803
3804 // We don't support switch statements inside loops.
3805 if (!isa<BranchInst>(BB->getTerminator())) {
3806 emitAnalysis(VectorizationReport(BB->getTerminator())
3807 << "loop contains a switch statement");
3808 return false;
3809 }
3810
3811 // We must be able to predicate all blocks that need to be predicated.
3812 if (blockNeedsPredication(BB)) {
3813 if (!blockCanBePredicated(BB, SafePointes)) {
3814 emitAnalysis(VectorizationReport(BB->getTerminator())
3815 << "control flow cannot be substituted for a select");
3816 return false;
3817 }
3818 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3819 emitAnalysis(VectorizationReport(BB->getTerminator())
3820 << "control flow cannot be substituted for a select");
3821 return false;
3822 }
3823 }
3824
3825 // We can if-convert this loop.
3826 return true;
3827}
3828
3829bool LoopVectorizationLegality::canVectorize() {
3830 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3831 // be canonicalized.
3832 if (!TheLoop->getLoopPreheader()) {
3833 emitAnalysis(
3834 VectorizationReport() <<
3835 "loop control flow is not understood by vectorizer");
3836 return false;
3837 }
3838
3839 // We can only vectorize innermost loops.
3840 if (!TheLoop->getSubLoopsVector().empty()) {
3841 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3842 return false;
3843 }
3844
3845 // We must have a single backedge.
3846 if (TheLoop->getNumBackEdges() != 1) {
3847 emitAnalysis(
3848 VectorizationReport() <<
3849 "loop control flow is not understood by vectorizer");
3850 return false;
3851 }
3852
3853 // We must have a single exiting block.
3854 if (!TheLoop->getExitingBlock()) {
3855 emitAnalysis(
3856 VectorizationReport() <<
3857 "loop control flow is not understood by vectorizer");
3858 return false;
3859 }
3860
3861 // We only handle bottom-tested loops, i.e. loop in which the condition is
3862 // checked at the end of each iteration. With that we can assume that all
3863 // instructions in the loop are executed the same number of times.
3864 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3865 emitAnalysis(
3866 VectorizationReport() <<
3867 "loop control flow is not understood by vectorizer");
3868 return false;
3869 }
3870
3871 // We need to have a loop header.
3872 DEBUG(dbgs() << "LV: Found a loop: " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop: " <<
TheLoop->getHeader()->getName() << '\n'; } } while
(0)
3873 TheLoop->getHeader()->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop: " <<
TheLoop->getHeader()->getName() << '\n'; } } while
(0);
3874
3875 // Check if we can if-convert non-single-bb loops.
3876 unsigned NumBlocks = TheLoop->getNumBlocks();
3877 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3878 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Can't if-convert the loop.\n"
; } } while (0);
3879 return false;
3880 }
3881
3882 // ScalarEvolution needs to be able to find the exit count.
3883 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3884 if (ExitCount == SE->getCouldNotCompute()) {
3885 emitAnalysis(VectorizationReport() <<
3886 "could not determine number of loop iterations");
3887 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: SCEV could not compute the loop exit count.\n"
; } } while (0);
3888 return false;
3889 }
3890
3891 // Check if we can vectorize the instructions and CFG in this loop.
3892 if (!canVectorizeInstrs()) {
3893 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Can't vectorize the instructions or CFG\n"
; } } while (0);
3894 return false;
3895 }
3896
3897 // Go over each instruction and look at memory deps.
3898 if (!canVectorizeMemory()) {
3899 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Can't vectorize due to memory conflicts\n"
; } } while (0);
3900 return false;
3901 }
3902
3903 // Collect all of the variables that remain uniform after vectorization.
3904 collectLoopUniforms();
3905
3906 DEBUG(dbgs() << "LV: We can vectorize this loop" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)"
: "") <<"!\n"; } } while (0)
3907 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)"
: "") <<"!\n"; } } while (0)
3908 "")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)"
: "") <<"!\n"; } } while (0)
3909 <<"!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)"
: "") <<"!\n"; } } while (0);
3910
3911 // Analyze interleaved memory accesses.
3912 if (EnableInterleavedMemAccesses)
3913 InterleaveInfo.analyzeInterleaving(Strides);
3914
3915 // Okay! We can vectorize. At this point we don't have any other mem analysis
3916 // which may limit our maximum vectorization factor, so just return true with
3917 // no restrictions.
3918 return true;
3919}
3920
3921static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3922 if (Ty->isPointerTy())
3923 return DL.getIntPtrType(Ty);
3924
3925 // It is possible that char's or short's overflow when we ask for the loop's
3926 // trip count, work around this by changing the type size.
3927 if (Ty->getScalarSizeInBits() < 32)
3928 return Type::getInt32Ty(Ty->getContext());
3929
3930 return Ty;
3931}
3932
3933static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3934 Ty0 = convertPointerToIntegerType(DL, Ty0);
3935 Ty1 = convertPointerToIntegerType(DL, Ty1);
3936 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3937 return Ty0;
3938 return Ty1;
3939}
3940
3941/// \brief Check that the instruction has outside loop users and is not an
3942/// identified reduction variable.
3943static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3944 SmallPtrSetImpl<Value *> &Reductions) {
3945 // Reduction instructions are allowed to have exit users. All other
3946 // instructions must not have external users.
3947 if (!Reductions.count(Inst))
3948 //Check that all of the users of the loop are inside the BB.
3949 for (User *U : Inst->users()) {
3950 Instruction *UI = cast<Instruction>(U);
3951 // This user may be a reduction exit value.
3952 if (!TheLoop->contains(UI)) {
3953 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an outside user for : "
<< *UI << '\n'; } } while (0);
3954 return true;
3955 }
3956 }
3957 return false;
3958}
3959
3960bool LoopVectorizationLegality::canVectorizeInstrs() {
3961 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3962 BasicBlock *Header = TheLoop->getHeader();
3963
3964 // Look for the attribute signaling the absence of NaNs.
3965 Function &F = *Header->getParent();
3966 const DataLayout &DL = F.getParent()->getDataLayout();
3967 if (F.hasFnAttribute("no-nans-fp-math"))
3968 HasFunNoNaNAttr =
3969 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3970
3971 // For each block in the loop.
3972 for (Loop::block_iterator bb = TheLoop->block_begin(),
3973 be = TheLoop->block_end(); bb != be; ++bb) {
3974
3975 // Scan the instructions in the block and look for hazards.
3976 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3977 ++it) {
3978
3979 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3980 Type *PhiTy = Phi->getType();
3981 // Check that this PHI type is allowed.
3982 if (!PhiTy->isIntegerTy() &&
3983 !PhiTy->isFloatingPointTy() &&
3984 !PhiTy->isPointerTy()) {
3985 emitAnalysis(VectorizationReport(it)
3986 << "loop control flow is not understood by vectorizer");
3987 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an non-int non-pointer PHI.\n"
; } } while (0);
3988 return false;
3989 }
3990
3991 // If this PHINode is not in the header block, then we know that we
3992 // can convert it to select during if-conversion. No need to check if
3993 // the PHIs in this block are induction or reduction variables.
3994 if (*bb != Header) {
3995 // Check that this instruction has no outside users or is an
3996 // identified reduction value with an outside user.
3997 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3998 continue;
3999 emitAnalysis(VectorizationReport(it) <<
4000 "value could not be identified as "
4001 "an induction or reduction variable");
4002 return false;
4003 }
4004
4005 // We only allow if-converted PHIs with exactly two incoming values.
4006 if (Phi->getNumIncomingValues() != 2) {
4007 emitAnalysis(VectorizationReport(it)
4008 << "control flow not understood by vectorizer");
4009 DEBUG(dbgs() << "LV: Found an invalid PHI.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an invalid PHI.\n"
; } } while (0);
4010 return false;
4011 }
4012
4013 // This is the value coming from the preheader.
4014 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
4015 ConstantInt *StepValue = nullptr;
4016 // Check if this is an induction variable.
4017 InductionKind IK = isInductionVariable(Phi, StepValue);
4018
4019 if (IK_NoInduction != IK) {
4020 // Get the widest type.
4021 if (!WidestIndTy)
4022 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4023 else
4024 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4025
4026 // Int inductions are special because we only allow one IV.
4027 if (IK == IK_IntInduction && StepValue->isOne()) {
4028 // Use the phi node with the widest type as induction. Use the last
4029 // one if there are multiple (no good reason for doing this other
4030 // than it is expedient).
4031 if (!Induction || PhiTy == WidestIndTy)
4032 Induction = Phi;
4033 }
4034
4035 DEBUG(dbgs() << "LV: Found an induction variable.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an induction variable.\n"
; } } while (0);
4036 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
4037
4038 // Until we explicitly handle the case of an induction variable with
4039 // an outside loop user we have to give up vectorizing this loop.
4040 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4041 emitAnalysis(VectorizationReport(it) <<
4042 "use of induction value outside of the "
4043 "loop is not handled by vectorizer");
4044 return false;
4045 }
4046
4047 continue;
4048 }
4049
4050 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4051 Reductions[Phi])) {
4052 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4053 continue;
4054 }
4055
4056 emitAnalysis(VectorizationReport(it) <<
4057 "value that could not be identified as "
4058 "reduction is used outside the loop");
4059 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an unidentified PHI."
<< *Phi <<"\n"; } } while (0);
4060 return false;
4061 }// end of PHI handling
4062
4063 // We handle calls that:
4064 // * Are debug info intrinsics.
4065 // * Have a mapping to an IR intrinsic.
4066 // * Have a vector version available.
4067 CallInst *CI = dyn_cast<CallInst>(it);
4068 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4069 !(CI->getCalledFunction() && TLI &&
4070 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4071 emitAnalysis(VectorizationReport(it) <<
4072 "call instruction cannot be vectorized");
4073 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n"
; } } while (0);
4074 return false;
4075 }
4076
4077 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4078 // second argument is the same (i.e. loop invariant)
4079 if (CI &&
4080 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4081 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4082 emitAnalysis(VectorizationReport(it)
4083 << "intrinsic instruction cannot be vectorized");
4084 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found unvectorizable intrinsic "
<< *CI << "\n"; } } while (0);
4085 return false;
4086 }
4087 }
4088
4089 // Check that the instruction return type is vectorizable.
4090 // Also, we can't vectorize extractelement instructions.
4091 if ((!VectorType::isValidElementType(it->getType()) &&
4092 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4093 emitAnalysis(VectorizationReport(it)
4094 << "instruction return type cannot be vectorized");
4095 DEBUG(dbgs() << "LV: Found unvectorizable type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found unvectorizable type.\n"
; } } while (0);
4096 return false;
4097 }
4098
4099 // Check that the stored type is vectorizable.
4100 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4101 Type *T = ST->getValueOperand()->getType();
4102 if (!VectorType::isValidElementType(T)) {
4103 emitAnalysis(VectorizationReport(ST) <<
4104 "store instruction cannot be vectorized");
4105 return false;
4106 }
4107 if (EnableMemAccessVersioning)
4108 collectStridedAccess(ST);
4109 }
4110
4111 if (EnableMemAccessVersioning)
4112 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4113 collectStridedAccess(LI);
4114
4115 // Reduction instructions are allowed to have exit users.
4116 // All other instructions must not have external users.
4117 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4118 emitAnalysis(VectorizationReport(it) <<
4119 "value cannot be used outside the loop");
4120 return false;
4121 }
4122
4123 } // next instr.
4124
4125 }
4126
4127 if (!Induction) {
4128 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Did not find one integer induction var.\n"
; } } while (0);
4129 if (Inductions.empty()) {
4130 emitAnalysis(VectorizationReport()
4131 << "loop induction variable could not be identified");
4132 return false;
4133 }
4134 }
4135
4136 return true;
4137}
4138
4139///\brief Remove GEPs whose indices but the last one are loop invariant and
4140/// return the induction operand of the gep pointer.
4141static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
4142 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
4143 if (!GEP)
4144 return Ptr;
4145
4146 unsigned InductionOperand = getGEPInductionOperand(GEP);
4147
4148 // Check that all of the gep indices are uniform except for our induction
4149 // operand.
4150 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
4151 if (i != InductionOperand &&
4152 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
4153 return Ptr;
4154 return GEP->getOperand(InductionOperand);
4155}
4156
4157///\brief Look for a cast use of the passed value.
4158static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
4159 Value *UniqueCast = nullptr;
4160 for (User *U : Ptr->users()) {
4161 CastInst *CI = dyn_cast<CastInst>(U);
4162 if (CI && CI->getType() == Ty) {
4163 if (!UniqueCast)
4164 UniqueCast = CI;
4165 else
4166 return nullptr;
4167 }
4168 }
4169 return UniqueCast;
4170}
4171
4172///\brief Get the stride of a pointer access in a loop.
4173/// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
4174/// pointer to the Value, or null otherwise.
4175static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
4176 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4177 if (!PtrTy || PtrTy->isAggregateType())
4178 return nullptr;
4179
4180 // Try to remove a gep instruction to make the pointer (actually index at this
4181 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
4182 // pointer, otherwise, we are analyzing the index.
4183 Value *OrigPtr = Ptr;
4184
4185 // The size of the pointer access.
4186 int64_t PtrAccessSize = 1;
4187
4188 Ptr = stripGetElementPtr(Ptr, SE, Lp);
4189 const SCEV *V = SE->getSCEV(Ptr);
4190
4191 if (Ptr != OrigPtr)
4192 // Strip off casts.
4193 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
4194 V = C->getOperand();
4195
4196 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
4197 if (!S)
4198 return nullptr;
4199
4200 V = S->getStepRecurrence(*SE);
4201 if (!V)
4202 return nullptr;
4203
4204 // Strip off the size of access multiplication if we are still analyzing the
4205 // pointer.
4206 if (OrigPtr == Ptr) {
4207 const DataLayout &DL = Lp->getHeader()->getModule()->getDataLayout();
4208 DL.getTypeAllocSize(PtrTy->getElementType());
4209 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
4210 if (M->getOperand(0)->getSCEVType() != scConstant)
4211 return nullptr;
4212
4213 const APInt &APStepVal =
4214 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
4215
4216 // Huge step value - give up.
4217 if (APStepVal.getBitWidth() > 64)
4218 return nullptr;
4219
4220 int64_t StepVal = APStepVal.getSExtValue();
4221 if (PtrAccessSize != StepVal)
4222 return nullptr;
4223 V = M->getOperand(1);
4224 }
4225 }
4226
4227 // Strip off casts.
4228 Type *StripedOffRecurrenceCast = nullptr;
4229 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
4230 StripedOffRecurrenceCast = C->getType();
4231 V = C->getOperand();
4232 }
4233
4234 // Look for the loop invariant symbolic value.
4235 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
4236 if (!U)
4237 return nullptr;
4238
4239 Value *Stride = U->getValue();
4240 if (!Lp->isLoopInvariant(Stride))
4241 return nullptr;
4242
4243 // If we have stripped off the recurrence cast we have to make sure that we
4244 // return the value that is used in this loop so that we can replace it later.
4245 if (StripedOffRecurrenceCast)
4246 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
4247
4248 return Stride;
4249}
4250
4251void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4252 Value *Ptr = nullptr;
4253 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4254 Ptr = LI->getPointerOperand();
4255 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4256 Ptr = SI->getPointerOperand();
4257 else
4258 return;
4259
4260 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4261 if (!Stride)
4262 return;
4263
4264 DEBUG(dbgs() << "LV: Found a strided access that we can version")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a strided access that we can version"
; } } while (0);
4265 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << " Ptr: " << *Ptr
<< " Stride: " << *Stride << "\n"; } } while
(0);
4266 Strides[Ptr] = Stride;
4267 StrideSet.insert(Stride);
4268}
4269
4270void LoopVectorizationLegality::collectLoopUniforms() {
4271 // We now know that the loop is vectorizable!
4272 // Collect variables that will remain uniform after vectorization.
4273 std::vector<Value*> Worklist;
4274 BasicBlock *Latch = TheLoop->getLoopLatch();
4275
4276 // Start with the conditional branch and walk up the block.
4277 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4278
4279 // Also add all consecutive pointer values; these values will be uniform
4280 // after vectorization (and subsequent cleanup) and, until revectorization is
4281 // supported, all dependencies must also be uniform.
4282 for (Loop::block_iterator B = TheLoop->block_begin(),
4283 BE = TheLoop->block_end(); B != BE; ++B)
4284 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4285 I != IE; ++I)
4286 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4287 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4288
4289 while (!Worklist.empty()) {
4290 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4291 Worklist.pop_back();
4292
4293 // Look at instructions inside this loop.
4294 // Stop when reaching PHI nodes.
4295 // TODO: we need to follow values all over the loop, not only in this block.
4296 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4297 continue;
4298
4299 // This is a known uniform.
4300 Uniforms.insert(I);
4301
4302 // Insert all operands.
4303 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4304 }
4305}
4306
4307bool LoopVectorizationLegality::canVectorizeMemory() {
4308 LAI = &LAA->getInfo(TheLoop, Strides);
4309 auto &OptionalReport = LAI->getReport();
4310 if (OptionalReport)
4311 emitAnalysis(VectorizationReport(*OptionalReport));
4312 if (!LAI->canVectorizeMemory())
4313 return false;
4314
4315 if (LAI->hasStoreToLoopInvariantAddress()) {
4316 emitAnalysis(
4317 VectorizationReport()
4318 << "write to a loop invariant address could not be vectorized");
4319 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We don't allow storing to uniform addresses\n"
; } } while (0);
4320 return false;
4321 }
4322
4323 if (LAI->getNumRuntimePointerChecks() >
4324 VectorizerParams::RuntimeMemoryCheckThreshold) {
4325 emitAnalysis(VectorizationReport()
4326 << LAI->getNumRuntimePointerChecks() << " exceeds limit of "
4327 << VectorizerParams::RuntimeMemoryCheckThreshold
4328 << " dependent memory operations checked at runtime");
4329 DEBUG(dbgs() << "LV: Too many memory checks needed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Too many memory checks needed.\n"
; } } while (0);
4330 return false;
4331 }
4332 return true;
4333}
4334
4335LoopVectorizationLegality::InductionKind
4336LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4337 ConstantInt *&StepValue) {
4338 if (!isInductionPHI(Phi, SE, StepValue))
4339 return IK_NoInduction;
4340
4341 Type *PhiTy = Phi->getType();
4342 // Found an Integer induction variable.
4343 if (PhiTy->isIntegerTy())
4344 return IK_IntInduction;
4345 // Found an Pointer induction variable.
4346 return IK_PtrInduction;
4347}
4348
4349bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4350 Value *In0 = const_cast<Value*>(V);
4351 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4352 if (!PN)
4353 return false;
4354
4355 return Inductions.count(PN);
4356}
4357
4358bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4359 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4360}
4361
4362bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4363 SmallPtrSetImpl<Value *> &SafePtrs) {
4364
4365 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4366 // Check that we don't have a constant expression that can trap as operand.
4367 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4368 OI != OE; ++OI) {
4369 if (Constant *C = dyn_cast<Constant>(*OI))
4370 if (C->canTrap())
4371 return false;
4372 }
4373 // We might be able to hoist the load.
4374 if (it->mayReadFromMemory()) {
4375 LoadInst *LI = dyn_cast<LoadInst>(it);
4376 if (!LI)
4377 return false;
4378 if (!SafePtrs.count(LI->getPointerOperand())) {
4379 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4380 MaskedOp.insert(LI);
4381 continue;
4382 }
4383 return false;
4384 }
4385 }
4386
4387 // We don't predicate stores at the moment.
4388 if (it->mayWriteToMemory()) {
4389 StoreInst *SI = dyn_cast<StoreInst>(it);
4390 // We only support predication of stores in basic blocks with one
4391 // predecessor.
4392 if (!SI)
4393 return false;
4394
4395 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4396 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4397
4398 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4399 !isSinglePredecessor) {
4400 // Build a masked store if it is legal for the target, otherwise scalarize
4401 // the block.
4402 bool isLegalMaskedOp =
4403 isLegalMaskedStore(SI->getValueOperand()->getType(),
4404 SI->getPointerOperand());
4405 if (isLegalMaskedOp) {
4406 --NumPredStores;
4407 MaskedOp.insert(SI);
4408 continue;
4409 }
4410 return false;
4411 }
4412 }
4413 if (it->mayThrow())
4414 return false;
4415
4416 // The instructions below can trap.
4417 switch (it->getOpcode()) {
4418 default: continue;
4419 case Instruction::UDiv:
4420 case Instruction::SDiv:
4421 case Instruction::URem:
4422 case Instruction::SRem:
4423 return false;
4424 }
4425 }
4426
4427 return true;
4428}
4429
4430void InterleavedAccessInfo::collectConstStridedAccesses(
4431 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4432 const ValueToValueMap &Strides) {
4433 // Holds load/store instructions in program order.
4434 SmallVector<Instruction *, 16> AccessList;
4435
4436 for (auto *BB : TheLoop->getBlocks()) {
4437 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4438
4439 for (auto &I : *BB) {
4440 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4441 continue;
4442 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4443 if (IsPred)
4444 return;
4445
4446 AccessList.push_back(&I);
4447 }
4448 }
4449
4450 if (AccessList.empty())
4451 return;
4452
4453 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4454 for (auto I : AccessList) {
4455 LoadInst *LI = dyn_cast<LoadInst>(I);
4456 StoreInst *SI = dyn_cast<StoreInst>(I);
4457
4458 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4459 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4460
4461 // The factor of the corresponding interleave group.
4462 unsigned Factor = std::abs(Stride);
4463
4464 // Ignore the access if the factor is too small or too large.
4465 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4466 continue;
4467
4468 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4469 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4470 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4471
4472 // An alignment of 0 means target ABI alignment.
4473 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4474 if (!Align)
4475 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4476
4477 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4478 }
4479}
4480
4481// Analyze interleaved accesses and collect them into interleave groups.
4482//
4483// Notice that the vectorization on interleaved groups will change instruction
4484// orders and may break dependences. But the memory dependence check guarantees
4485// that there is no overlap between two pointers of different strides, element
4486// sizes or underlying bases.
4487//
4488// For pointers sharing the same stride, element size and underlying base, no
4489// need to worry about Read-After-Write dependences and Write-After-Read
4490// dependences.
4491//
4492// E.g. The RAW dependence: A[i] = a;
4493// b = A[i];
4494// This won't exist as it is a store-load forwarding conflict, which has
4495// already been checked and forbidden in the dependence check.
4496//
4497// E.g. The WAR dependence: a = A[i]; // (1)
4498// A[i] = b; // (2)
4499// The store group of (2) is always inserted at or below (2), and the load group
4500// of (1) is always inserted at or above (1). The dependence is safe.
4501void InterleavedAccessInfo::analyzeInterleaving(
4502 const ValueToValueMap &Strides) {
4503 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Analyzing interleaved accesses...\n"
; } } while (0);
4504
4505 // Holds all the stride accesses.
4506 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4507 collectConstStridedAccesses(StrideAccesses, Strides);
4508
4509 if (StrideAccesses.empty())
4510 return;
4511
4512 // Holds all interleaved store groups temporarily.
4513 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4514
4515 // Search the load-load/write-write pair B-A in bottom-up order and try to
4516 // insert B into the interleave group of A according to 3 rules:
4517 // 1. A and B have the same stride.
4518 // 2. A and B have the same memory object size.
4519 // 3. B belongs to the group according to the distance.
4520 //
4521 // The bottom-up order can avoid breaking the Write-After-Write dependences
4522 // between two pointers of the same base.
4523 // E.g. A[i] = a; (1)
4524 // A[i] = b; (2)
4525 // A[i+1] = c (3)
4526 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4527 // above (1), which guarantees that (1) is always above (2).
4528 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4529 ++I) {
4530 Instruction *A = I->first;
4531 StrideDescriptor DesA = I->second;
4532
4533 InterleaveGroup *Group = getInterleaveGroup(A);
4534 if (!Group) {
4535 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Creating an interleave group with:"
<< *A << '\n'; } } while (0);
4536 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4537 }
4538
4539 if (A->mayWriteToMemory())
4540 StoreGroups.insert(Group);
4541
4542 for (auto II = std::next(I); II != E; ++II) {
4543 Instruction *B = II->first;
4544 StrideDescriptor DesB = II->second;
4545
4546 // Ignore if B is already in a group or B is a different memory operation.
4547 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4548 continue;
4549
4550 // Check the rule 1 and 2.
4551 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4552 continue;
4553
4554 // Calculate the distance and prepare for the rule 3.
4555 const SCEVConstant *DistToA =
4556 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4557 if (!DistToA)
4558 continue;
4559
4560 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4561
4562 // Skip if the distance is not multiple of size as they are not in the
4563 // same group.
4564 if (DistanceToA % static_cast<int>(DesA.Size))
4565 continue;
4566
4567 // The index of B is the index of A plus the related index to A.
4568 int IndexB =
4569 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4570
4571 // Try to insert B into the group.
4572 if (Group->insertMember(B, IndexB, DesB.Align)) {
4573 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Inserted:" <<
*B << '\n' << " into the interleave group with"
<< *A << '\n'; } } while (0)
4574 << " into the interleave group with" << *A << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Inserted:" <<
*B << '\n' << " into the interleave group with"
<< *A << '\n'; } } while (0);
4575 InterleaveGroupMap[B] = Group;
4576
4577 // Set the first load in program order as the insert position.
4578 if (B->mayReadFromMemory())
4579 Group->setInsertPos(B);
4580 }
4581 } // Iteration on instruction B
4582 } // Iteration on instruction A
4583
4584 // Remove interleaved store groups with gaps.
4585 for (InterleaveGroup *Group : StoreGroups)
4586 if (Group->getNumMembers() != Group->getFactor())
4587 releaseGroup(Group);
4588}
4589
4590LoopVectorizationCostModel::VectorizationFactor
4591LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4592 // Width 1 means no vectorize
4593 VectorizationFactor Factor = { 1U, 0U };
4594 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4595 emitAnalysis(VectorizationReport() <<
4596 "runtime pointer checks needed. Enable vectorization of this "
4597 "loop with '#pragma clang loop vectorize(enable)' when "
4598 "compiling with -Os");
4599 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"
; } } while (0);
4600 return Factor;
4601 }
4602
4603 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4604 emitAnalysis(VectorizationReport() <<
4605 "store that is conditionally executed prevents vectorization");
4606 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: No vectorization. There are conditional stores.\n"
; } } while (0);
4607 return Factor;
4608 }
4609
4610 // Find the trip count.
4611 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4612 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found trip count: "
<< TC << '\n'; } } while (0);
4613
4614 unsigned WidestType = getWidestType();
4615 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4616 unsigned MaxSafeDepDist = -1U;
4617 if (Legal->getMaxSafeDepDistBytes() != -1U)
4618 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4619 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4620 WidestRegister : MaxSafeDepDist);
4621 unsigned MaxVectorSize = WidestRegister / WidestType;
4622 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest type: "
<< WidestType << " bits.\n"; } } while (0);
4623 DEBUG(dbgs() << "LV: The Widest register is: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest register is: "
<< WidestRegister << " bits.\n"; } } while (0)
4624 << WidestRegister << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest register is: "
<< WidestRegister << " bits.\n"; } } while (0);
4625
4626 if (MaxVectorSize == 0) {
4627 DEBUG(dbgs() << "LV: The target has no vector registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has no vector registers.\n"
; } } while (0);
4628 MaxVectorSize = 1;
4629 }
4630
4631 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"((MaxVectorSize <= 64 && "Did not expect to pack so many elements"
" into one vector!") ? static_cast<void> (0) : __assert_fail
("MaxVectorSize <= 64 && \"Did not expect to pack so many elements\" \" into one vector!\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4632, __PRETTY_FUNCTION__))
4632 " into one vector!")((MaxVectorSize <= 64 && "Did not expect to pack so many elements"
" into one vector!") ? static_cast<void> (0) : __assert_fail
("MaxVectorSize <= 64 && \"Did not expect to pack so many elements\" \" into one vector!\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4632, __PRETTY_FUNCTION__));
4633
4634 unsigned VF = MaxVectorSize;
4635
4636 // If we optimize the program for size, avoid creating the tail loop.
4637 if (OptForSize) {
4638 // If we are unable to calculate the trip count then don't try to vectorize.
4639 if (TC < 2) {
4640 emitAnalysis
4641 (VectorizationReport() <<
4642 "unable to calculate the loop count due to complex control flow");
4643 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. A tail loop is required in Os.\n"
; } } while (0);
4644 return Factor;
4645 }
4646
4647 // Find the maximum SIMD width that can fit within the trip count.
4648 VF = TC % MaxVectorSize;
4649
4650 if (VF == 0)
4651 VF = MaxVectorSize;
4652 else {
4653 // If the trip count that we found modulo the vectorization factor is not
4654 // zero then we require a tail.
4655 emitAnalysis(VectorizationReport() <<
4656 "cannot optimize for size and vectorize at the "
4657 "same time. Enable vectorization of this loop "
4658 "with '#pragma clang loop vectorize(enable)' "
4659 "when compiling with -Os");
4660 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. A tail loop is required in Os.\n"
; } } while (0);
4661 return Factor;
4662 }
4663 }
4664
4665 int UserVF = Hints->getWidth();
4666 if (UserVF != 0) {
4667 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two")((isPowerOf2_32(UserVF) && "VF needs to be a power of two"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(UserVF) && \"VF needs to be a power of two\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4667, __PRETTY_FUNCTION__));
4668 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Using user VF " <<
UserVF << ".\n"; } } while (0);
4669
4670 Factor.Width = UserVF;
4671 return Factor;
4672 }
4673
4674 float Cost = expectedCost(1);
4675#ifndef NDEBUG
4676 const float ScalarCost = Cost;
4677#endif /* NDEBUG */
4678 unsigned Width = 1;
4679 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalar loop costs: "
<< (int)ScalarCost << ".\n"; } } while (0);
4680
4681 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4682 // Ignore scalar width, because the user explicitly wants vectorization.
4683 if (ForceVectorization && VF > 1) {
4684 Width = 2;
4685 Cost = expectedCost(Width) / (float)Width;
4686 }
4687
4688 for (unsigned i=2; i <= VF; i*=2) {
4689 // Notice that the vector loop needs to be executed less times, so
4690 // we need to divide the cost of the vector loops by the width of
4691 // the vector elements.
4692 float VectorCost = expectedCost(i) / (float)i;
4693 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Vector loop of width "
<< i << " costs: " << (int)VectorCost <<
".\n"; } } while (0)
4694 (int)VectorCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Vector loop of width "
<< i << " costs: " << (int)VectorCost <<
".\n"; } } while (0);
4695 if (VectorCost < Cost) {
4696 Cost = VectorCost;
4697 Width = i;
4698 }
4699 }
4700
4701 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { if (ForceVectorization && Width
> 1 && Cost >= ScalarCost) dbgs() << "LV: Vectorization seems to be not beneficial, "
<< "but was forced by a user.\n"; } } while (0)
4702 << "LV: Vectorization seems to be not beneficial, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { if (ForceVectorization && Width
> 1 && Cost >= ScalarCost) dbgs() << "LV: Vectorization seems to be not beneficial, "
<< "but was forced by a user.\n"; } } while (0)
4703 << "but was forced by a user.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { if (ForceVectorization && Width
> 1 && Cost >= ScalarCost) dbgs() << "LV: Vectorization seems to be not beneficial, "
<< "but was forced by a user.\n"; } } while (0);
4704 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Selecting VF: "<<
Width << ".\n"; } } while (0);
4705 Factor.Width = Width;
4706 Factor.Cost = Width * Cost;
4707 return Factor;
4708}
4709
4710unsigned LoopVectorizationCostModel::getWidestType() {
4711 unsigned MaxWidth = 8;
4712 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4713
4714 // For each block.
4715 for (Loop::block_iterator bb = TheLoop->block_begin(),
4716 be = TheLoop->block_end(); bb != be; ++bb) {
4717 BasicBlock *BB = *bb;
4718
4719 // For each instruction in the loop.
4720 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4721 Type *T = it->getType();
4722
4723 // Ignore ephemeral values.
4724 if (EphValues.count(it))
4725 continue;
4726
4727 // Only examine Loads, Stores and PHINodes.
4728 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4729 continue;
4730
4731 // Examine PHI nodes that are reduction variables.
4732 if (PHINode *PN = dyn_cast<PHINode>(it))
4733 if (!Legal->getReductionVars()->count(PN))
4734 continue;
4735
4736 // Examine the stored values.
4737 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4738 T = ST->getValueOperand()->getType();
4739
4740 // Ignore loaded pointer types and stored pointer types that are not
4741 // consecutive. However, we do want to take consecutive stores/loads of
4742 // pointer vectors into account.
4743 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4744 continue;
4745
4746 MaxWidth = std::max(MaxWidth,
4747 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4748 }
4749 }
4750
4751 return MaxWidth;
4752}
4753
4754unsigned
4755LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4756 unsigned VF,
4757 unsigned LoopCost) {
4758
4759 // -- The unroll heuristics --
4760 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4761 // There are many micro-architectural considerations that we can't predict
4762 // at this level. For example, frontend pressure (on decode or fetch) due to
4763 // code size, or the number and capabilities of the execution ports.
4764 //
4765 // We use the following heuristics to select the unroll factor:
4766 // 1. If the code has reductions, then we unroll in order to break the cross
4767 // iteration dependency.
4768 // 2. If the loop is really small, then we unroll in order to reduce the loop
4769 // overhead.
4770 // 3. We don't unroll if we think that we will spill registers to memory due
4771 // to the increased register pressure.
4772
4773 // Use the user preference, unless 'auto' is selected.
4774 int UserUF = Hints->getInterleave();
4775 if (UserUF != 0)
1
Assuming 'UserUF' is equal to 0
→
2
←
Taking false branch
→
4776 return UserUF;
4777
4778 // When we optimize for size, we don't unroll.
4779 if (OptForSize)
3
←
Assuming 'OptForSize' is 0
→
4
←
Taking false branch
→
4780 return 1;
4781
4782 // We used the distance for the unroll factor.
4783 if (Legal->getMaxSafeDepDistBytes() != -1U)
5
←
Taking false branch
→
4784 return 1;
4785
4786 // Do not unroll loops with a relatively small trip count.
4787 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4788 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
6
←
Assuming 'TC' is <= 1
→
4789 return 1;
4790
4791 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
7
←
Assuming 'VF' is <= 1
→
4792 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (0)
4793 " registers\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (0);
4794
4795 if (VF == 1) {
8
←
Assuming 'VF' is not equal to 1
→
9
←
Taking false branch
→
4796 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4797 TargetNumRegisters = ForceTargetNumScalarRegs;
4798 } else {
4799 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
10
←
Taking false branch
→
4800 TargetNumRegisters = ForceTargetNumVectorRegs;
4801 }
4802
4803 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4804 // We divide by these constants so assume that we have at least one
4805 // instruction that uses at least one register.
4806 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4807 R.NumInstructions = std::max(R.NumInstructions, 1U);
4808
4809 // We calculate the unroll factor using the following formula.
4810 // Subtract the number of loop invariants from the number of available
4811 // registers. These registers are used by all of the unrolled instances.
4812 // Next, divide the remaining registers by the number of registers that is
4813 // required by the loop, in order to estimate how many parallel instances
4814 // fit without causing spills. All of this is rounded down if necessary to be
4815 // a power of two. We want power of two unroll factors to simplify any
4816 // addressing operations or alignment considerations.
4817 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4818 R.MaxLocalUsers);
4819
4820 // Don't count the induction variable as unrolled.
4821 if (EnableIndVarRegisterHeur)
11
←
Taking false branch
→
4822 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4823 std::max(1U, (R.MaxLocalUsers - 1)));
4824
4825 // Clamp the unroll factor ranges to reasonable factors.
4826 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor(VF);
4827
4828 // Check if the user has overridden the unroll max.
4829 if (VF == 1) {
12
←
Taking false branch
→
4830 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4831 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4832 } else {
4833 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
13
←
Taking false branch
→
4834 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4835 }
4836
4837 // If we did not calculate the cost for VF (because the user selected the VF)
4838 // then we calculate the cost of VF here.
4839 if (LoopCost == 0)
14
←
Assuming 'LoopCost' is equal to 0
→
15
←
Taking true branch
→
4840 LoopCost = expectedCost(VF);
16
←
Calling 'LoopVectorizationCostModel::expectedCost'
→
22
←
Returning from 'LoopVectorizationCostModel::expectedCost'
→
23
←
The value 0 is assigned to 'LoopCost'
→
4841
4842 // Clamp the calculated UF to be between the 1 and the max unroll factor
4843 // that the target allows.
4844 if (UF > MaxInterleaveSize)
24
←
Taking false branch
→
4845 UF = MaxInterleaveSize;
4846 else if (UF < 1)
25
←
Assuming 'UF' is >= 1
→
26
←
Taking false branch
→
4847 UF = 1;
4848
4849 // Unroll if we vectorized this loop and there is a reduction that could
4850 // benefit from unrolling.
4851 if (VF > 1 && Legal->getReductionVars()->size()) {
4852 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Unrolling because of reductions.\n"
; } } while (0);
4853 return UF;
4854 }
4855
4856 // Note that if we've already vectorized the loop we will have done the
4857 // runtime check and so unrolling won't require further checks.
4858 bool UnrollingRequiresRuntimePointerCheck =
4859 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4860
4861 // We want to unroll small loops in order to reduce the loop overhead and
4862 // potentially expose ILP opportunities.
4863 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop cost is " <<
LoopCost << '\n'; } } while (0);
4864 if (!UnrollingRequiresRuntimePointerCheck &&
27
←
Taking true branch
→
4865 LoopCost < SmallLoopCost) {
4866 // We assume that the cost overhead is 1 and we use the cost model
4867 // to estimate the cost of the loop and unroll until the cost of the
4868 // loop overhead is about 5% of the cost of the loop.
4869 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
28
←
Division by zero
4870
4871 // Unroll until store/load ports (estimated by max unroll factor) are
4872 // saturated.
4873 unsigned NumStores = Legal->getNumStores();
4874 unsigned NumLoads = Legal->getNumLoads();
4875 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4876 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4877
4878 // If we have a scalar reduction (vector reductions are already dealt with
4879 // by this point), we can increase the critical path length if the loop
4880 // we're unrolling is inside another loop. Limit, by default to 2, so the
4881 // critical path only gets increased by one reduction operation.
4882 if (Legal->getReductionVars()->size() &&
4883 TheLoop->getLoopDepth() > 1) {
4884 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4885 SmallUF = std::min(SmallUF, F);
4886 StoresUF = std::min(StoresUF, F);
4887 LoadsUF = std::min(LoadsUF, F);
4888 }
4889
4890 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4891 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Unrolling to saturate store or load ports.\n"
; } } while (0);
4892 return std::max(StoresUF, LoadsUF);
4893 }
4894
4895 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Unrolling to reduce branch cost.\n"
; } } while (0);
4896 return SmallUF;
4897 }
4898
4899 // Unroll if this is a large loop (small loops are already dealt with by this
4900 // point) that could benefit from interleaved unrolling.
4901 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4902 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4903 DEBUG(dbgs() << "LV: Unrolling to expose ILP.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Unrolling to expose ILP.\n"
; } } while (0);
4904 return UF;
4905 }
4906
4907 DEBUG(dbgs() << "LV: Not Unrolling.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not Unrolling.\n";
} } while (0);
4908 return 1;
4909}
4910
4911LoopVectorizationCostModel::RegisterUsage
4912LoopVectorizationCostModel::calculateRegisterUsage() {
4913 // This function calculates the register usage by measuring the highest number
4914 // of values that are alive at a single location. Obviously, this is a very
4915 // rough estimation. We scan the loop in a topological order in order and
4916 // assign a number to each instruction. We use RPO to ensure that defs are
4917 // met before their users. We assume that each instruction that has in-loop
4918 // users starts an interval. We record every time that an in-loop value is
4919 // used, so we have a list of the first and last occurrences of each
4920 // instruction. Next, we transpose this data structure into a multi map that
4921 // holds the list of intervals that *end* at a specific location. This multi
4922 // map allows us to perform a linear search. We scan the instructions linearly
4923 // and record each time that a new interval starts, by placing it in a set.
4924 // If we find this value in the multi-map then we remove it from the set.
4925 // The max register usage is the maximum size of the set.
4926 // We also search for instructions that are defined outside the loop, but are
4927 // used inside the loop. We need this number separately from the max-interval
4928 // usage number because when we unroll, loop-invariant values do not take
4929 // more register.
4930 LoopBlocksDFS DFS(TheLoop);
4931 DFS.perform(LI);
4932
4933 RegisterUsage R;
4934 R.NumInstructions = 0;
4935
4936 // Each 'key' in the map opens a new interval. The values
4937 // of the map are the index of the 'last seen' usage of the
4938 // instruction that is the key.
4939 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4940 // Maps instruction to its index.
4941 DenseMap<unsigned, Instruction*> IdxToInstr;
4942 // Marks the end of each interval.
4943 IntervalMap EndPoint;
4944 // Saves the list of instruction indices that are used in the loop.
4945 SmallSet<Instruction*, 8> Ends;
4946 // Saves the list of values that are used in the loop but are
4947 // defined outside the loop, such as arguments and constants.
4948 SmallPtrSet<Value*, 8> LoopInvariants;
4949
4950 unsigned Index = 0;
4951 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4952 be = DFS.endRPO(); bb != be; ++bb) {
4953 R.NumInstructions += (*bb)->size();
4954 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4955 ++it) {
4956 Instruction *I = it;
4957 IdxToInstr[Index++] = I;
4958
4959 // Save the end location of each USE.
4960 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4961 Value *U = I->getOperand(i);
4962 Instruction *Instr = dyn_cast<Instruction>(U);
4963
4964 // Ignore non-instruction values such as arguments, constants, etc.
4965 if (!Instr) continue;
4966
4967 // If this instruction is outside the loop then record it and continue.
4968 if (!TheLoop->contains(Instr)) {
4969 LoopInvariants.insert(Instr);
4970 continue;
4971 }
4972
4973 // Overwrite previous end points.
4974 EndPoint[Instr] = Index;
4975 Ends.insert(Instr);
4976 }
4977 }
4978 }
4979
4980 // Saves the list of intervals that end with the index in 'key'.
4981 typedef SmallVector<Instruction*, 2> InstrList;
4982 DenseMap<unsigned, InstrList> TransposeEnds;
4983
4984 // Transpose the EndPoints to a list of values that end at each index.
4985 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4986 it != e; ++it)
4987 TransposeEnds[it->second].push_back(it->first);
4988
4989 SmallSet<Instruction*, 8> OpenIntervals;
4990 unsigned MaxUsage = 0;
4991
4992
4993 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): Calculating max register usage:\n"
; } } while (0);
4994 for (unsigned int i = 0; i < Index; ++i) {
4995 Instruction *I = IdxToInstr[i];
4996 // Ignore instructions that are never used within the loop.
4997 if (!Ends.count(I)) continue;
4998
4999 // Ignore ephemeral values.
5000 if (EphValues.count(I))
5001 continue;
5002
5003 // Remove all of the instructions that end at this location.
5004 InstrList &List = TransposeEnds[i];
5005 for (unsigned int j=0, e = List.size(); j < e; ++j)
5006 OpenIntervals.erase(List[j]);
5007
5008 // Count the number of live interals.
5009 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5010
5011 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): At #" <<
i << " Interval # " << OpenIntervals.size() <<
'\n'; } } while (0)
5012 OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): At #" <<
i << " Interval # " << OpenIntervals.size() <<
'\n'; } } while (0);
5013
5014 // Add the current instruction to the list of open intervals.
5015 OpenIntervals.insert(I);
5016 }
5017
5018 unsigned Invariant = LoopInvariants.size();
5019 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): Found max usage: "
<< MaxUsage << '\n'; } } while (0);
5020 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): Found invariant usage: "
<< Invariant << '\n'; } } while (0);
5021 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): LoopSize: " <<
R.NumInstructions << '\n'; } } while (0);
5022
5023 R.LoopInvariantRegs = Invariant;
5024 R.MaxLocalUsers = MaxUsage;
5025 return R;
5026}
5027
5028unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5029 unsigned Cost = 0;
17
←
'Cost' initialized to 0
→
5030
5031 // For each block.
5032 for (Loop::block_iterator bb = TheLoop->block_begin(),
20
←
Loop condition is false. Execution continues on line 5067
→
5033 be = TheLoop->block_end(); bb != be; ++bb) {
18
←
Calling 'operator!='
→
19
←
Returning from 'operator!='
→
5034 unsigned BlockCost = 0;
5035 BasicBlock *BB = *bb;
5036
5037 // For each instruction in the old loop.
5038 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5039 // Skip dbg intrinsics.
5040 if (isa<DbgInfoIntrinsic>(it))
5041 continue;
5042
5043 // Ignore ephemeral values.
5044 if (EphValues.count(it))
5045 continue;
5046
5047 unsigned C = getInstructionCost(it, VF);
5048
5049 // Check if we should override the cost.
5050 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5051 C = ForceTargetInstructionCost;
5052
5053 BlockCost += C;
5054 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of "
<< C << " for VF " << VF << " For instruction: "
<< *it << '\n'; } } while (0)
5055 VF << " For instruction: " << *it << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of "
<< C << " for VF " << VF << " For instruction: "
<< *it << '\n'; } } while (0);
5056 }
5057
5058 // We assume that if-converted blocks have a 50% chance of being executed.
5059 // When the code is scalar then some of the blocks are avoided due to CF.
5060 // When the code is vectorized we execute all code paths.
5061 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5062 BlockCost /= 2;
5063
5064 Cost += BlockCost;
5065 }
5066
5067 return Cost;
21
←
Returning zero (loaded from 'Cost')
→
5068}
5069
5070/// \brief Check whether the address computation for a non-consecutive memory
5071/// access looks like an unlikely candidate for being merged into the indexing
5072/// mode.
5073///
5074/// We look for a GEP which has one index that is an induction variable and all
5075/// other indices are loop invariant. If the stride of this access is also
5076/// within a small bound we decide that this address computation can likely be
5077/// merged into the addressing mode.
5078/// In all other cases, we identify the address computation as complex.
5079static bool isLikelyComplexAddressComputation(Value *Ptr,
5080 LoopVectorizationLegality *Legal,
5081 ScalarEvolution *SE,
5082 const Loop *TheLoop) {
5083 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5084 if (!Gep)
5085 return true;
5086
5087 // We are looking for a gep with all loop invariant indices except for one
5088 // which should be an induction variable.
5089 unsigned NumOperands = Gep->getNumOperands();
5090 for (unsigned i = 1; i < NumOperands; ++i) {
5091 Value *Opd = Gep->getOperand(i);
5092 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5093 !Legal->isInductionVariable(Opd))
5094 return true;
5095 }
5096
5097 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5098 // can likely be merged into the address computation.
5099 unsigned MaxMergeDistance = 64;
5100
5101 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5102 if (!AddRec)
5103 return true;
5104
5105 // Check the step is constant.
5106 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5107 // Calculate the pointer stride and check if it is consecutive.
5108 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5109 if (!C)
5110 return true;
5111
5112 const APInt &APStepVal = C->getValue()->getValue();
5113
5114 // Huge step value - give up.
5115 if (APStepVal.getBitWidth() > 64)
5116 return true;
5117
5118 int64_t StepVal = APStepVal.getSExtValue();
5119
5120 return StepVal > MaxMergeDistance;
5121}
5122
5123static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5124 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5125 return true;
5126 return false;
5127}
5128
5129unsigned
5130LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5131 // If we know that this instruction will remain uniform, check the cost of
5132 // the scalar version.
5133 if (Legal->isUniformAfterVectorization(I))
5134 VF = 1;
5135
5136 Type *RetTy = I->getType();
5137 Type *VectorTy = ToVectorTy(RetTy, VF);
5138
5139 // TODO: We need to estimate the cost of intrinsic calls.
5140 switch (I->getOpcode()) {
5141 case Instruction::GetElementPtr:
5142 // We mark this instruction as zero-cost because the cost of GEPs in
5143 // vectorized code depends on whether the corresponding memory instruction
5144 // is scalarized or not. Therefore, we handle GEPs with the memory
5145 // instruction cost.
5146 return 0;
5147 case Instruction::Br: {
5148 return TTI.getCFInstrCost(I->getOpcode());
5149 }
5150 case Instruction::PHI:
5151 //TODO: IF-converted IFs become selects.
5152 return 0;
5153 case Instruction::Add:
5154 case Instruction::FAdd:
5155 case Instruction::Sub:
5156 case Instruction::FSub:
5157 case Instruction::Mul:
5158 case Instruction::FMul:
5159 case Instruction::UDiv:
5160 case Instruction::SDiv:
5161 case Instruction::FDiv:
5162 case Instruction::URem:
5163 case Instruction::SRem:
5164 case Instruction::FRem:
5165 case Instruction::Shl:
5166 case Instruction::LShr:
5167 case Instruction::AShr:
5168 case Instruction::And:
5169 case Instruction::Or:
5170 case Instruction::Xor: {
5171 // Since we will replace the stride by 1 the multiplication should go away.
5172 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5173 return 0;
5174 // Certain instructions can be cheaper to vectorize if they have a constant
5175 // second vector operand. One example of this are shifts on x86.
5176 TargetTransformInfo::OperandValueKind Op1VK =
5177 TargetTransformInfo::OK_AnyValue;
5178 TargetTransformInfo::OperandValueKind Op2VK =
5179 TargetTransformInfo::OK_AnyValue;
5180 TargetTransformInfo::OperandValueProperties Op1VP =
5181 TargetTransformInfo::OP_None;
5182 TargetTransformInfo::OperandValueProperties Op2VP =
5183 TargetTransformInfo::OP_None;
5184 Value *Op2 = I->getOperand(1);
5185
5186 // Check for a splat of a constant or for a non uniform vector of constants.
5187 if (isa<ConstantInt>(Op2)) {
5188 ConstantInt *CInt = cast<ConstantInt>(Op2);
5189 if (CInt && CInt->getValue().isPowerOf2())
5190 Op2VP = TargetTransformInfo::OP_PowerOf2;
5191 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5192 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5193 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5194 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5195 if (SplatValue) {
5196 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5197 if (CInt && CInt->getValue().isPowerOf2())
5198 Op2VP = TargetTransformInfo::OP_PowerOf2;
5199 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5200 }
5201 }
5202
5203 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5204 Op1VP, Op2VP);
5205 }
5206 case Instruction::Select: {
5207 SelectInst *SI = cast<SelectInst>(I);
5208 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5209 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5210 Type *CondTy = SI->getCondition()->getType();
5211 if (!ScalarCond)
5212 CondTy = VectorType::get(CondTy, VF);
5213
5214 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5215 }
5216 case Instruction::ICmp:
5217 case Instruction::FCmp: {
5218 Type *ValTy = I->getOperand(0)->getType();
5219 VectorTy = ToVectorTy(ValTy, VF);
5220 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5221 }
5222 case Instruction::Store:
5223 case Instruction::Load: {
5224 StoreInst *SI = dyn_cast<StoreInst>(I);
5225 LoadInst *LI = dyn_cast<LoadInst>(I);
5226 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5227 LI->getType());
5228 VectorTy = ToVectorTy(ValTy, VF);
5229
5230 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5231 unsigned AS = SI ? SI->getPointerAddressSpace() :
5232 LI->getPointerAddressSpace();
5233 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5234 // We add the cost of address computation here instead of with the gep
5235 // instruction because only here we know whether the operation is
5236 // scalarized.
5237 if (VF == 1)
5238 return TTI.getAddressComputationCost(VectorTy) +
5239 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5240
5241 // For an interleaved access, calculate the total cost of the whole
5242 // interleave group.
5243 if (Legal->isAccessInterleaved(I)) {
5244 auto Group = Legal->getInterleavedAccessGroup(I);
5245 assert(Group && "Fail to get an interleaved access group.")((Group && "Fail to get an interleaved access group."
) ? static_cast<void> (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5245, __PRETTY_FUNCTION__));
5246
5247 // Only calculate the cost once at the insert position.
5248 if (Group->getInsertPos() != I)
5249 return 0;
5250
5251 unsigned InterleaveFactor = Group->getFactor();
5252 Type *WideVecTy =
5253 VectorType::get(VectorTy->getVectorElementType(),
5254 VectorTy->getVectorNumElements() * InterleaveFactor);
5255
5256 // Holds the indices of existing members in an interleaved load group.
5257 // An interleaved store group doesn't need this as it dones't allow gaps.
5258 SmallVector<unsigned, 4> Indices;
5259 if (LI) {
5260 for (unsigned i = 0; i < InterleaveFactor; i++)
5261 if (Group->getMember(i))
5262 Indices.push_back(i);
5263 }
5264
5265 // Calculate the cost of the whole interleaved group.
5266 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5267 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5268 Group->getAlignment(), AS);
5269
5270 if (Group->isReverse())
5271 Cost +=
5272 Group->getNumMembers() *
5273 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5274
5275 // FIXME: The interleaved load group with a huge gap could be even more
5276 // expensive than scalar operations. Then we could ignore such group and
5277 // use scalar operations instead.
5278 return Cost;
5279 }
5280
5281 // Scalarized loads/stores.
5282 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5283 bool Reverse = ConsecutiveStride < 0;
5284 const DataLayout &DL = I->getModule()->getDataLayout();
5285 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5286 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5287 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5288 bool IsComplexComputation =
5289 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5290 unsigned Cost = 0;
5291 // The cost of extracting from the value vector and pointer vector.
5292 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5293 for (unsigned i = 0; i < VF; ++i) {
5294 // The cost of extracting the pointer operand.
5295 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5296 // In case of STORE, the cost of ExtractElement from the vector.
5297 // In case of LOAD, the cost of InsertElement into the returned
5298 // vector.
5299 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5300 Instruction::InsertElement,
5301 VectorTy, i);
5302 }
5303
5304 // The cost of the scalar loads/stores.
5305 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5306 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5307 Alignment, AS);
5308 return Cost;
5309 }
5310
5311 // Wide load/stores.
5312 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5313 if (Legal->isMaskRequired(I))
5314 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5315 AS);
5316 else
5317 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5318
5319 if (Reverse)
5320 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5321 VectorTy, 0);
5322 return Cost;
5323 }
5324 case Instruction::ZExt:
5325 case Instruction::SExt:
5326 case Instruction::FPToUI:
5327 case Instruction::FPToSI:
5328 case Instruction::FPExt:
5329 case Instruction::PtrToInt:
5330 case Instruction::IntToPtr:
5331 case Instruction::SIToFP:
5332 case Instruction::UIToFP:
5333 case Instruction::Trunc:
5334 case Instruction::FPTrunc:
5335 case Instruction::BitCast: {
5336 // We optimize the truncation of induction variable.
5337 // The cost of these is the same as the scalar operation.
5338 if (I->getOpcode() == Instruction::Trunc &&
5339 Legal->isInductionVariable(I->getOperand(0)))
5340 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5341 I->getOperand(0)->getType());
5342
5343 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5344 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5345 }
5346 case Instruction::Call: {
5347 bool NeedToScalarize;
5348 CallInst *CI = cast<CallInst>(I);
5349 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5350 if (getIntrinsicIDForCall(CI, TLI))
5351 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5352 return CallCost;
5353 }
5354 default: {
5355 // We are scalarizing the instruction. Return the cost of the scalar
5356 // instruction, plus the cost of insert and extract into vector
5357 // elements, times the vector width.
5358 unsigned Cost = 0;
5359
5360 if (!RetTy->isVoidTy() && VF != 1) {
5361 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5362 VectorTy);
5363 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5364 VectorTy);
5365
5366 // The cost of inserting the results plus extracting each one of the
5367 // operands.
5368 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5369 }
5370
5371 // The cost of executing VF copies of the scalar instruction. This opcode
5372 // is unknown. Assume that it is the same as 'mul'.
5373 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5374 return Cost;
5375 }
5376 }// end of switch.
5377}
5378
5379char LoopVectorize::ID = 0;
5380static const char lv_name[] = "Loop Vectorization";
5381INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)static void* initializeLoopVectorizePassOnce(PassRegistry &
Registry) {
5382INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
5383INITIALIZE_AG_DEPENDENCY(AliasAnalysis)initializeAliasAnalysisAnalysisGroup(Registry);
5384INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
5385INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)initializeBlockFrequencyInfoPass(Registry);
5386INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
5387INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)initializeScalarEvolutionPass(Registry);
5388INITIALIZE_PASS_DEPENDENCY(LCSSA)initializeLCSSAPass(Registry);
5389INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
5390INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
5391INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)initializeLoopAccessAnalysisPass(Registry);
5392INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo(lv_name, "loop-vectorize", & LoopVectorize
::ID, PassInfo::NormalCtor_t(callDefaultCtor< LoopVectorize
>), false, false); Registry.registerPass(*PI, true); return
PI; } void llvm::initializeLoopVectorizePass(PassRegistry &
Registry) { static volatile sys::cas_flag initialized = 0; sys
::cas_flag old_val = sys::CompareAndSwap(&initialized, 1,
0); if (old_val == 0) { initializeLoopVectorizePassOnce(Registry
); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5392); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5392, &initialized); initialized = 2; AnnotateIgnoreWritesEnd
("/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5392); } else { sys::cas_flag tmp = initialized; sys::MemoryFence
(); while (tmp != 2) { tmp = initialized; sys::MemoryFence();
} } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5392, &initialized); }
5393
5394namespace llvm {
5395 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5396 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5397 }
5398}
5399
5400bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5401 // Check for a store.
5402 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5403 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5404
5405 // Check for a load.
5406 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5407 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5408
5409 return false;
5410}
5411
5412
5413void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5414 bool IfPredicateStore) {
5415 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors")((!Instr->getType()->isAggregateType() && "Can't handle vectors"
) ? static_cast<void> (0) : __assert_fail ("!Instr->getType()->isAggregateType() && \"Can't handle vectors\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5415, __PRETTY_FUNCTION__));
5416 // Holds vector parameters or scalars, in case of uniform vals.
5417 SmallVector<VectorParts, 4> Params;
5418
5419 setDebugLocFromInst(Builder, Instr);
5420
5421 // Find all of the vectorized parameters.
5422 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5423 Value *SrcOp = Instr->getOperand(op);
5424
5425 // If we are accessing the old induction variable, use the new one.
5426 if (SrcOp == OldInduction) {
5427 Params.push_back(getVectorValue(SrcOp));
5428 continue;
5429 }
5430
5431 // Try using previously calculated values.
5432 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5433
5434 // If the src is an instruction that appeared earlier in the basic block
5435 // then it should already be vectorized.
5436 if (SrcInst && OrigLoop->contains(SrcInst)) {
5437 assert(WidenMap.has(SrcInst) && "Source operand is unavailable")((WidenMap.has(SrcInst) && "Source operand is unavailable"
) ? static_cast<void> (0) : __assert_fail ("WidenMap.has(SrcInst) && \"Source operand is unavailable\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5437, __PRETTY_FUNCTION__));
5438 // The parameter is a vector value from earlier.
5439 Params.push_back(WidenMap.get(SrcInst));
5440 } else {
5441 // The parameter is a scalar from outside the loop. Maybe even a constant.
5442 VectorParts Scalars;
5443 Scalars.append(UF, SrcOp);
5444 Params.push_back(Scalars);
5445 }
5446 }
5447
5448 assert(Params.size() == Instr->getNumOperands() &&((Params.size() == Instr->getNumOperands() && "Invalid number of operands"
) ? static_cast<void> (0) : __assert_fail ("Params.size() == Instr->getNumOperands() && \"Invalid number of operands\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5449, __PRETTY_FUNCTION__))
5449 "Invalid number of operands")((Params.size() == Instr->getNumOperands() && "Invalid number of operands"
) ? static_cast<void> (0) : __assert_fail ("Params.size() == Instr->getNumOperands() && \"Invalid number of operands\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5449, __PRETTY_FUNCTION__));
5450
5451 // Does this instruction return a value ?
5452 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5453
5454 Value *UndefVec = IsVoidRetTy ? nullptr :
5455 UndefValue::get(Instr->getType());
5456 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5457 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5458
5459 Instruction *InsertPt = Builder.GetInsertPoint();
5460 BasicBlock *IfBlock = Builder.GetInsertBlock();
5461 BasicBlock *CondBlock = nullptr;
5462
5463 VectorParts Cond;
5464 Loop *VectorLp = nullptr;
5465 if (IfPredicateStore) {
5466 assert(Instr->getParent()->getSinglePredecessor() &&((Instr->getParent()->getSinglePredecessor() &&
"Only support single predecessor blocks") ? static_cast<void
> (0) : __assert_fail ("Instr->getParent()->getSinglePredecessor() && \"Only support single predecessor blocks\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5467, __PRETTY_FUNCTION__))
5467 "Only support single predecessor blocks")((Instr->getParent()->getSinglePredecessor() &&
"Only support single predecessor blocks") ? static_cast<void
> (0) : __assert_fail ("Instr->getParent()->getSinglePredecessor() && \"Only support single predecessor blocks\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5467, __PRETTY_FUNCTION__));
5468 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5469 Instr->getParent());
5470 VectorLp = LI->getLoopFor(IfBlock);
5471 assert(VectorLp && "Must have a loop for this block")((VectorLp && "Must have a loop for this block") ? static_cast
<void> (0) : __assert_fail ("VectorLp && \"Must have a loop for this block\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5471, __PRETTY_FUNCTION__));
5472 }
5473
5474 // For each vector unroll 'part':
5475 for (unsigned Part = 0; Part < UF; ++Part) {
5476 // For each scalar that we create:
5477
5478 // Start an "if (pred) a[i] = ..." block.
5479 Value *Cmp = nullptr;
5480 if (IfPredicateStore) {
5481 if (Cond[Part]->getType()->isVectorTy())
5482 Cond[Part] =
5483 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5484 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5485 ConstantInt::get(Cond[Part]->getType(), 1));
5486 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5487 LoopVectorBody.push_back(CondBlock);
5488 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5489 // Update Builder with newly created basic block.
5490 Builder.SetInsertPoint(InsertPt);
5491 }
5492
5493 Instruction *Cloned = Instr->clone();
5494 if (!IsVoidRetTy)
5495 Cloned->setName(Instr->getName() + ".cloned");
5496 // Replace the operands of the cloned instructions with extracted scalars.
5497 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5498 Value *Op = Params[op][Part];
5499 Cloned->setOperand(op, Op);
5500 }
5501
5502 // Place the cloned scalar in the new loop.
5503 Builder.Insert(Cloned);
5504
5505 // If the original scalar returns a value we need to place it in a vector
5506 // so that future users will be able to use it.
5507 if (!IsVoidRetTy)
5508 VecResults[Part] = Cloned;
5509
5510 // End if-block.
5511 if (IfPredicateStore) {
5512 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5513 LoopVectorBody.push_back(NewIfBlock);
5514 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5515 Builder.SetInsertPoint(InsertPt);
5516 Instruction *OldBr = IfBlock->getTerminator();
5517 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5518 OldBr->eraseFromParent();
5519 IfBlock = NewIfBlock;
5520 }
5521 }
5522}
5523
5524void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5525 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5526 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5527
5528 return scalarizeInstruction(Instr, IfPredicateStore);
5529}
5530
5531Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5532 return Vec;
5533}
5534
5535Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5536 return V;
5537}
5538
5539Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5540 // When unrolling and the VF is 1, we only need to add a simple scalar.
5541 Type *ITy = Val->getType();
5542 assert(!ITy->isVectorTy() && "Val must be a scalar")((!ITy->isVectorTy() && "Val must be a scalar") ? static_cast
<void> (0) : __assert_fail ("!ITy->isVectorTy() && \"Val must be a scalar\""
, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn240924/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5542, __PRETTY_FUNCTION__));
5543 Constant *C = ConstantInt::get(ITy, StartIdx);
5544 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");
5545}