Bug Summary

File:lib/Transforms/Vectorize/LoopVectorize.cpp
Warning:line 6095, column 11
Called C++ object pointer is null

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/LoopVectorize.h"
50#include "llvm/ADT/DenseMap.h"
51#include "llvm/ADT/Hashing.h"
52#include "llvm/ADT/MapVector.h"
53#include "llvm/ADT/Optional.h"
54#include "llvm/ADT/SCCIterator.h"
55#include "llvm/ADT/SetVector.h"
56#include "llvm/ADT/SmallPtrSet.h"
57#include "llvm/ADT/SmallSet.h"
58#include "llvm/ADT/SmallVector.h"
59#include "llvm/ADT/Statistic.h"
60#include "llvm/ADT/StringExtras.h"
61#include "llvm/Analysis/CodeMetrics.h"
62#include "llvm/Analysis/GlobalsModRef.h"
63#include "llvm/Analysis/LoopInfo.h"
64#include "llvm/Analysis/LoopIterator.h"
65#include "llvm/Analysis/LoopPass.h"
66#include "llvm/Analysis/ScalarEvolutionExpander.h"
67#include "llvm/Analysis/ScalarEvolutionExpressions.h"
68#include "llvm/Analysis/ValueTracking.h"
69#include "llvm/Analysis/VectorUtils.h"
70#include "llvm/IR/Constants.h"
71#include "llvm/IR/DataLayout.h"
72#include "llvm/IR/DebugInfo.h"
73#include "llvm/IR/DerivedTypes.h"
74#include "llvm/IR/DiagnosticInfo.h"
75#include "llvm/IR/Dominators.h"
76#include "llvm/IR/Function.h"
77#include "llvm/IR/IRBuilder.h"
78#include "llvm/IR/Instructions.h"
79#include "llvm/IR/IntrinsicInst.h"
80#include "llvm/IR/LLVMContext.h"
81#include "llvm/IR/Module.h"
82#include "llvm/IR/PatternMatch.h"
83#include "llvm/IR/Type.h"
84#include "llvm/IR/User.h"
85#include "llvm/IR/Value.h"
86#include "llvm/IR/ValueHandle.h"
87#include "llvm/IR/Verifier.h"
88#include "llvm/Pass.h"
89#include "llvm/Support/BranchProbability.h"
90#include "llvm/Support/CommandLine.h"
91#include "llvm/Support/Debug.h"
92#include "llvm/Support/raw_ostream.h"
93#include "llvm/Transforms/Scalar.h"
94#include "llvm/Transforms/Utils/BasicBlockUtils.h"
95#include "llvm/Transforms/Utils/Local.h"
96#include "llvm/Transforms/Utils/LoopSimplify.h"
97#include "llvm/Transforms/Utils/LoopUtils.h"
98#include "llvm/Transforms/Utils/LoopVersioning.h"
99#include "llvm/Transforms/Vectorize.h"
100#include <algorithm>
101#include <map>
102#include <tuple>
103
104using namespace llvm;
105using namespace llvm::PatternMatch;
106
107#define LV_NAME"loop-vectorize" "loop-vectorize"
108#define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize"
109
110STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized"
, "Number of loops vectorized", {0}, false}
;
111STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed"
, "Number of loops analyzed for vectorization", {0}, false}
;
112
113static cl::opt<bool>
114 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
115 cl::desc("Enable if-conversion during vectorization."));
116
117/// We don't vectorize loops with a known constant trip count below this number.
118static cl::opt<unsigned> TinyTripCountVectorThreshold(
119 "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
120 cl::desc("Don't vectorize loops with a constant "
121 "trip count that is smaller than this "
122 "value."));
123
124static cl::opt<bool> MaximizeBandwidth(
125 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
126 cl::desc("Maximize bandwidth when selecting vectorization factor which "
127 "will be determined by the smallest type in loop."));
128
129static cl::opt<bool> EnableInterleavedMemAccesses(
130 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
131 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
132
133/// Maximum factor for an interleaved memory access.
134static cl::opt<unsigned> MaxInterleaveGroupFactor(
135 "max-interleave-group-factor", cl::Hidden,
136 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
137 cl::init(8));
138
139/// We don't interleave loops with a known constant trip count below this
140/// number.
141static const unsigned TinyTripCountInterleaveThreshold = 128;
142
143static cl::opt<unsigned> ForceTargetNumScalarRegs(
144 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
145 cl::desc("A flag that overrides the target's number of scalar registers."));
146
147static cl::opt<unsigned> ForceTargetNumVectorRegs(
148 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
149 cl::desc("A flag that overrides the target's number of vector registers."));
150
151/// Maximum vectorization interleave count.
152static const unsigned MaxInterleaveFactor = 16;
153
154static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
155 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's max interleave factor for "
157 "scalar loops."));
158
159static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
160 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's max interleave factor for "
162 "vectorized loops."));
163
164static cl::opt<unsigned> ForceTargetInstructionCost(
165 "force-target-instruction-cost", cl::init(0), cl::Hidden,
166 cl::desc("A flag that overrides the target's expected cost for "
167 "an instruction to a single constant value. Mostly "
168 "useful for getting consistent testing."));
169
170static cl::opt<unsigned> SmallLoopCost(
171 "small-loop-cost", cl::init(20), cl::Hidden,
172 cl::desc(
173 "The cost of a loop that is considered 'small' by the interleaver."));
174
175static cl::opt<bool> LoopVectorizeWithBlockFrequency(
176 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
177 cl::desc("Enable the use of the block frequency analysis to access PGO "
178 "heuristics minimizing code growth in cold regions and being more "
179 "aggressive in hot regions."));
180
181// Runtime interleave loops for load/store throughput.
182static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
183 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
184 cl::desc(
185 "Enable runtime interleaving until load/store ports are saturated"));
186
187/// The number of stores in a loop that are allowed to need predication.
188static cl::opt<unsigned> NumberOfStoresToPredicate(
189 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
190 cl::desc("Max number of stores to be predicated behind an if."));
191
192static cl::opt<bool> EnableIndVarRegisterHeur(
193 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
194 cl::desc("Count the induction variable only once when interleaving"));
195
196static cl::opt<bool> EnableCondStoresVectorization(
197 "enable-cond-stores-vec", cl::init(true), cl::Hidden,
198 cl::desc("Enable if predication of stores during vectorization."));
199
200static cl::opt<unsigned> MaxNestedScalarReductionIC(
201 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
202 cl::desc("The maximum interleave count to use when interleaving a scalar "
203 "reduction in a nested loop."));
204
205static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
206 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
207 cl::desc("The maximum allowed number of runtime memory checks with a "
208 "vectorize(enable) pragma."));
209
210static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
211 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
212 cl::desc("The maximum number of SCEV checks allowed."));
213
214static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
215 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
216 cl::desc("The maximum number of SCEV checks allowed with a "
217 "vectorize(enable) pragma"));
218
219/// Create an analysis remark that explains why vectorization failed
220///
221/// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
222/// RemarkName is the identifier for the remark. If \p I is passed it is an
223/// instruction that prevents vectorization. Otherwise \p TheLoop is used for
224/// the location of the remark. \return the remark object that can be
225/// streamed to.
226static OptimizationRemarkAnalysis
227createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
228 Instruction *I = nullptr) {
229 Value *CodeRegion = TheLoop->getHeader();
230 DebugLoc DL = TheLoop->getStartLoc();
231
232 if (I) {
233 CodeRegion = I->getParent();
234 // If there is no debug location attached to the instruction, revert back to
235 // using the loop's.
236 if (I->getDebugLoc())
237 DL = I->getDebugLoc();
238 }
239
240 OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
241 R << "loop not vectorized: ";
242 return R;
243}
244
245namespace {
246
247// Forward declarations.
248class LoopVectorizeHints;
249class LoopVectorizationLegality;
250class LoopVectorizationCostModel;
251class LoopVectorizationRequirements;
252
253/// Returns true if the given loop body has a cycle, excluding the loop
254/// itself.
255static bool hasCyclesInLoopBody(const Loop &L) {
256 if (!L.empty())
257 return true;
258
259 for (const auto &SCC :
260 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
261 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
262 if (SCC.size() > 1) {
263 DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LVL: Detected a cycle in the loop body:\n"
; } } while (false)
;
264 DEBUG(L.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { L.dump(); } } while (false)
;
265 return true;
266 }
267 }
268 return false;
269}
270
271/// A helper function for converting Scalar types to vector types.
272/// If the incoming type is void, we return void. If the VF is 1, we return
273/// the scalar type.
274static Type *ToVectorTy(Type *Scalar, unsigned VF) {
275 if (Scalar->isVoidTy() || VF == 1)
276 return Scalar;
277 return VectorType::get(Scalar, VF);
278}
279
280/// A helper function that returns GEP instruction and knows to skip a
281/// 'bitcast'. The 'bitcast' may be skipped if the source and the destination
282/// pointee types of the 'bitcast' have the same size.
283/// For example:
284/// bitcast double** %var to i64* - can be skipped
285/// bitcast double** %var to i8* - can not
286static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
287
288 if (isa<GetElementPtrInst>(Ptr))
289 return cast<GetElementPtrInst>(Ptr);
290
291 if (isa<BitCastInst>(Ptr) &&
292 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) {
293 Type *BitcastTy = Ptr->getType();
294 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy();
295 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy))
296 return nullptr;
297 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType();
298 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType();
299 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout();
300 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty))
301 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
302 }
303 return nullptr;
304}
305
306// FIXME: The following helper functions have multiple implementations
307// in the project. They can be effectively organized in a common Load/Store
308// utilities unit.
309
310/// A helper function that returns the pointer operand of a load or store
311/// instruction.
312static Value *getPointerOperand(Value *I) {
313 if (auto *LI = dyn_cast<LoadInst>(I))
314 return LI->getPointerOperand();
315 if (auto *SI = dyn_cast<StoreInst>(I))
316 return SI->getPointerOperand();
317 return nullptr;
318}
319
320/// A helper function that returns the type of loaded or stored value.
321static Type *getMemInstValueType(Value *I) {
322 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
"Expected Load or Store instruction") ? static_cast<void>
(0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 323, __PRETTY_FUNCTION__))
323 "Expected Load or Store instruction")(((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
"Expected Load or Store instruction") ? static_cast<void>
(0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 323, __PRETTY_FUNCTION__))
;
324 if (auto *LI = dyn_cast<LoadInst>(I))
325 return LI->getType();
326 return cast<StoreInst>(I)->getValueOperand()->getType();
327}
328
329/// A helper function that returns the alignment of load or store instruction.
330static unsigned getMemInstAlignment(Value *I) {
331 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
"Expected Load or Store instruction") ? static_cast<void>
(0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 332, __PRETTY_FUNCTION__))
332 "Expected Load or Store instruction")(((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
"Expected Load or Store instruction") ? static_cast<void>
(0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 332, __PRETTY_FUNCTION__))
;
333 if (auto *LI = dyn_cast<LoadInst>(I))
334 return LI->getAlignment();
335 return cast<StoreInst>(I)->getAlignment();
336}
337
338/// A helper function that returns the address space of the pointer operand of
339/// load or store instruction.
340static unsigned getMemInstAddressSpace(Value *I) {
341 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
"Expected Load or Store instruction") ? static_cast<void>
(0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 342, __PRETTY_FUNCTION__))
342 "Expected Load or Store instruction")(((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
"Expected Load or Store instruction") ? static_cast<void>
(0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 342, __PRETTY_FUNCTION__))
;
343 if (auto *LI = dyn_cast<LoadInst>(I))
344 return LI->getPointerAddressSpace();
345 return cast<StoreInst>(I)->getPointerAddressSpace();
346}
347
348/// A helper function that returns true if the given type is irregular. The
349/// type is irregular if its allocated size doesn't equal the store size of an
350/// element of the corresponding vector type at the given vectorization factor.
351static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
352
353 // Determine if an array of VF elements of type Ty is "bitcast compatible"
354 // with a <VF x Ty> vector.
355 if (VF > 1) {
356 auto *VectorTy = VectorType::get(Ty, VF);
357 return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
358 }
359
360 // If the vectorization factor is one, we just check if an array of type Ty
361 // requires padding between elements.
362 return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
363}
364
365/// A helper function that returns the reciprocal of the block probability of
366/// predicated blocks. If we return X, we are assuming the predicated block
367/// will execute once for for every X iterations of the loop header.
368///
369/// TODO: We should use actual block probability here, if available. Currently,
370/// we always assume predicated blocks have a 50% chance of executing.
371static unsigned getReciprocalPredBlockProb() { return 2; }
372
373/// A helper function that adds a 'fast' flag to floating-point operations.
374static Value *addFastMathFlag(Value *V) {
375 if (isa<FPMathOperator>(V)) {
376 FastMathFlags Flags;
377 Flags.setUnsafeAlgebra();
378 cast<Instruction>(V)->setFastMathFlags(Flags);
379 }
380 return V;
381}
382
383/// A helper function that returns an integer or floating-point constant with
384/// value C.
385static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) {
386 return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C)
387 : ConstantFP::get(Ty, C);
388}
389
390/// InnerLoopVectorizer vectorizes loops which contain only one basic
391/// block to a specified vectorization factor (VF).
392/// This class performs the widening of scalars into vectors, or multiple
393/// scalars. This class also implements the following features:
394/// * It inserts an epilogue loop for handling loops that don't have iteration
395/// counts that are known to be a multiple of the vectorization factor.
396/// * It handles the code generation for reduction variables.
397/// * Scalarization (implementation using scalars) of un-vectorizable
398/// instructions.
399/// InnerLoopVectorizer does not perform any vectorization-legality
400/// checks, and relies on the caller to check for the different legality
401/// aspects. The InnerLoopVectorizer relies on the
402/// LoopVectorizationLegality class to provide information about the induction
403/// and reduction variables that were found to a given vectorization factor.
404class InnerLoopVectorizer {
405public:
406 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
407 LoopInfo *LI, DominatorTree *DT,
408 const TargetLibraryInfo *TLI,
409 const TargetTransformInfo *TTI, AssumptionCache *AC,
410 OptimizationRemarkEmitter *ORE, unsigned VecWidth,
411 unsigned UnrollFactor, LoopVectorizationLegality *LVL,
412 LoopVectorizationCostModel *CM)
413 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
414 AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
415 Builder(PSE.getSE()->getContext()), Induction(nullptr),
416 OldInduction(nullptr), VectorLoopValueMap(UnrollFactor, VecWidth),
417 TripCount(nullptr), VectorTripCount(nullptr), Legal(LVL), Cost(CM),
418 AddedSafetyChecks(false) {}
419
420 // Perform the actual loop widening (vectorization).
421 void vectorize() {
422 // Create a new empty loop. Unlink the old loop and connect the new one.
423 createEmptyLoop();
424 // Widen each instruction in the old loop to a new one in the new loop.
425 vectorizeLoop();
426 }
427
428 // Return true if any runtime check is added.
429 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
430
431 virtual ~InnerLoopVectorizer() {}
432
433protected:
434 /// A small list of PHINodes.
435 typedef SmallVector<PHINode *, 4> PhiVector;
436
437 /// A type for vectorized values in the new loop. Each value from the
438 /// original loop, when vectorized, is represented by UF vector values in the
439 /// new unrolled loop, where UF is the unroll factor.
440 typedef SmallVector<Value *, 2> VectorParts;
441
442 /// A type for scalarized values in the new loop. Each value from the
443 /// original loop, when scalarized, is represented by UF x VF scalar values
444 /// in the new unrolled loop, where UF is the unroll factor and VF is the
445 /// vectorization factor.
446 typedef SmallVector<SmallVector<Value *, 4>, 2> ScalarParts;
447
448 // When we if-convert we need to create edge masks. We have to cache values
449 // so that we don't end up with exponential recursion/IR.
450 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
451 EdgeMaskCache;
452
453 /// Create an empty loop, based on the loop ranges of the old loop.
454 void createEmptyLoop();
455
456 /// Set up the values of the IVs correctly when exiting the vector loop.
457 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
458 Value *CountRoundDown, Value *EndValue,
459 BasicBlock *MiddleBlock);
460
461 /// Create a new induction variable inside L.
462 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
463 Value *Step, Instruction *DL);
464 /// Copy and widen the instructions from the old loop.
465 virtual void vectorizeLoop();
466
467 /// Handle all cross-iteration phis in the header.
468 void fixCrossIterationPHIs();
469
470 /// Fix a first-order recurrence. This is the second phase of vectorizing
471 /// this phi node.
472 void fixFirstOrderRecurrence(PHINode *Phi);
473
474 /// Fix a reduction cross-iteration phi. This is the second phase of
475 /// vectorizing this phi node.
476 void fixReduction(PHINode *Phi);
477
478 /// \brief The Loop exit block may have single value PHI nodes where the
479 /// incoming value is 'Undef'. While vectorizing we only handled real values
480 /// that were defined inside the loop. Here we fix the 'undef case'.
481 /// See PR14725.
482 void fixLCSSAPHIs();
483
484 /// Iteratively sink the scalarized operands of a predicated instruction into
485 /// the block that was created for it.
486 void sinkScalarOperands(Instruction *PredInst);
487
488 /// Predicate conditional instructions that require predication on their
489 /// respective conditions.
490 void predicateInstructions();
491
492 /// Collect the instructions from the original loop that would be trivially
493 /// dead in the vectorized loop if generated.
494 void collectTriviallyDeadInstructions();
495
496 /// Shrinks vector element sizes to the smallest bitwidth they can be legally
497 /// represented as.
498 void truncateToMinimalBitwidths();
499
500 /// A helper function that computes the predicate of the block BB, assuming
501 /// that the header block of the loop is set to True. It returns the *entry*
502 /// mask for the block BB.
503 VectorParts createBlockInMask(BasicBlock *BB);
504 /// A helper function that computes the predicate of the edge between SRC
505 /// and DST.
506 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
507
508 /// A helper function to vectorize a single BB within the innermost loop.
509 void vectorizeBlockInLoop(BasicBlock *BB);
510
511 /// Vectorize a single PHINode in a block. This method handles the induction
512 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
513 /// arbitrary length vectors.
514 void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF);
515
516 /// Insert the new loop to the loop hierarchy and pass manager
517 /// and update the analysis passes.
518 void updateAnalysis();
519
520 /// This instruction is un-vectorizable. Implement it as a sequence
521 /// of scalars. If \p IfPredicateInstr is true we need to 'hide' each
522 /// scalarized instruction behind an if block predicated on the control
523 /// dependence of the instruction.
524 virtual void scalarizeInstruction(Instruction *Instr,
525 bool IfPredicateInstr = false);
526
527 /// Vectorize Load and Store instructions,
528 virtual void vectorizeMemoryInstruction(Instruction *Instr);
529
530 /// Create a broadcast instruction. This method generates a broadcast
531 /// instruction (shuffle) for loop invariant values and for the induction
532 /// value. If this is the induction variable then we extend it to N, N+1, ...
533 /// this is needed because each iteration in the loop corresponds to a SIMD
534 /// element.
535 virtual Value *getBroadcastInstrs(Value *V);
536
537 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
538 /// to each vector element of Val. The sequence starts at StartIndex.
539 /// \p Opcode is relevant for FP induction variable.
540 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
541 Instruction::BinaryOps Opcode =
542 Instruction::BinaryOpsEnd);
543
544 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
545 /// variable on which to base the steps, \p Step is the size of the step, and
546 /// \p EntryVal is the value from the original loop that maps to the steps.
547 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
548 /// can be a truncate instruction).
549 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal,
550 const InductionDescriptor &ID);
551
552 /// Create a vector induction phi node based on an existing scalar one. \p
553 /// EntryVal is the value from the original loop that maps to the vector phi
554 /// node, and \p Step is the loop-invariant step. If \p EntryVal is a
555 /// truncate instruction, instead of widening the original IV, we widen a
556 /// version of the IV truncated to \p EntryVal's type.
557 void createVectorIntOrFpInductionPHI(const InductionDescriptor &II,
558 Value *Step, Instruction *EntryVal);
559
560 /// Widen an integer or floating-point induction variable \p IV. If \p Trunc
561 /// is provided, the integer induction variable will first be truncated to
562 /// the corresponding type.
563 void widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc = nullptr);
564
565 /// Returns true if an instruction \p I should be scalarized instead of
566 /// vectorized for the chosen vectorization factor.
567 bool shouldScalarizeInstruction(Instruction *I) const;
568
569 /// Returns true if we should generate a scalar version of \p IV.
570 bool needsScalarInduction(Instruction *IV) const;
571
572 /// Return a constant reference to the VectorParts corresponding to \p V from
573 /// the original loop. If the value has already been vectorized, the
574 /// corresponding vector entry in VectorLoopValueMap is returned. If,
575 /// however, the value has a scalar entry in VectorLoopValueMap, we construct
576 /// new vector values on-demand by inserting the scalar values into vectors
577 /// with an insertelement sequence. If the value has been neither vectorized
578 /// nor scalarized, it must be loop invariant, so we simply broadcast the
579 /// value into vectors.
580 const VectorParts &getVectorValue(Value *V);
581
582 /// Return a value in the new loop corresponding to \p V from the original
583 /// loop at unroll index \p Part and vector index \p Lane. If the value has
584 /// been vectorized but not scalarized, the necessary extractelement
585 /// instruction will be generated.
586 Value *getScalarValue(Value *V, unsigned Part, unsigned Lane);
587
588 /// Try to vectorize the interleaved access group that \p Instr belongs to.
589 void vectorizeInterleaveGroup(Instruction *Instr);
590
591 /// Generate a shuffle sequence that will reverse the vector Vec.
592 virtual Value *reverseVector(Value *Vec);
593
594 /// Returns (and creates if needed) the original loop trip count.
595 Value *getOrCreateTripCount(Loop *NewLoop);
596
597 /// Returns (and creates if needed) the trip count of the widened loop.
598 Value *getOrCreateVectorTripCount(Loop *NewLoop);
599
600 /// Emit a bypass check to see if the trip count would overflow, or we
601 /// wouldn't have enough iterations to execute one vector loop.
602 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
603 /// Emit a bypass check to see if the vector trip count is nonzero.
604 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
605 /// Emit a bypass check to see if all of the SCEV assumptions we've
606 /// had to make are correct.
607 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
608 /// Emit bypass checks to check any memory assumptions we may have made.
609 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
610
611 /// Add additional metadata to \p To that was not present on \p Orig.
612 ///
613 /// Currently this is used to add the noalias annotations based on the
614 /// inserted memchecks. Use this for instructions that are *cloned* into the
615 /// vector loop.
616 void addNewMetadata(Instruction *To, const Instruction *Orig);
617
618 /// Add metadata from one instruction to another.
619 ///
620 /// This includes both the original MDs from \p From and additional ones (\see
621 /// addNewMetadata). Use this for *newly created* instructions in the vector
622 /// loop.
623 void addMetadata(Instruction *To, Instruction *From);
624
625 /// \brief Similar to the previous function but it adds the metadata to a
626 /// vector of instructions.
627 void addMetadata(ArrayRef<Value *> To, Instruction *From);
628
629 /// \brief Set the debug location in the builder using the debug location in
630 /// the instruction.
631 void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr);
632
633 /// This is a helper class for maintaining vectorization state. It's used for
634 /// mapping values from the original loop to their corresponding values in
635 /// the new loop. Two mappings are maintained: one for vectorized values and
636 /// one for scalarized values. Vectorized values are represented with UF
637 /// vector values in the new loop, and scalarized values are represented with
638 /// UF x VF scalar values in the new loop. UF and VF are the unroll and
639 /// vectorization factors, respectively.
640 ///
641 /// Entries can be added to either map with initVector and initScalar, which
642 /// initialize and return a constant reference to the new entry. If a
643 /// non-constant reference to a vector entry is required, getVector can be
644 /// used to retrieve a mutable entry. We currently directly modify the mapped
645 /// values during "fix-up" operations that occur once the first phase of
646 /// widening is complete. These operations include type truncation and the
647 /// second phase of recurrence widening.
648 ///
649 /// Otherwise, entries from either map should be accessed using the
650 /// getVectorValue or getScalarValue functions from InnerLoopVectorizer.
651 /// getVectorValue and getScalarValue coordinate to generate a vector or
652 /// scalar value on-demand if one is not yet available. When vectorizing a
653 /// loop, we visit the definition of an instruction before its uses. When
654 /// visiting the definition, we either vectorize or scalarize the
655 /// instruction, creating an entry for it in the corresponding map. (In some
656 /// cases, such as induction variables, we will create both vector and scalar
657 /// entries.) Then, as we encounter uses of the definition, we derive values
658 /// for each scalar or vector use unless such a value is already available.
659 /// For example, if we scalarize a definition and one of its uses is vector,
660 /// we build the required vector on-demand with an insertelement sequence
661 /// when visiting the use. Otherwise, if the use is scalar, we can use the
662 /// existing scalar definition.
663 struct ValueMap {
664
665 /// Construct an empty map with the given unroll and vectorization factors.
666 ValueMap(unsigned UnrollFactor, unsigned VecWidth)
667 : UF(UnrollFactor), VF(VecWidth) {
668 // The unroll and vectorization factors are only used in asserts builds
669 // to verify map entries are sized appropriately.
670 (void)UF;
671 (void)VF;
672 }
673
674 /// \return True if the map has a vector entry for \p Key.
675 bool hasVector(Value *Key) const { return VectorMapStorage.count(Key); }
676
677 /// \return True if the map has a scalar entry for \p Key.
678 bool hasScalar(Value *Key) const { return ScalarMapStorage.count(Key); }
679
680 /// \brief Map \p Key to the given VectorParts \p Entry, and return a
681 /// constant reference to the new vector map entry. The given key should
682 /// not already be in the map, and the given VectorParts should be
683 /// correctly sized for the current unroll factor.
684 const VectorParts &initVector(Value *Key, const VectorParts &Entry) {
685 assert(!hasVector(Key) && "Vector entry already initialized")((!hasVector(Key) && "Vector entry already initialized"
) ? static_cast<void> (0) : __assert_fail ("!hasVector(Key) && \"Vector entry already initialized\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 685, __PRETTY_FUNCTION__))
;
686 assert(Entry.size() == UF && "VectorParts has wrong dimensions")((Entry.size() == UF && "VectorParts has wrong dimensions"
) ? static_cast<void> (0) : __assert_fail ("Entry.size() == UF && \"VectorParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 686, __PRETTY_FUNCTION__))
;
687 VectorMapStorage[Key] = Entry;
688 return VectorMapStorage[Key];
689 }
690
691 /// \brief Map \p Key to the given ScalarParts \p Entry, and return a
692 /// constant reference to the new scalar map entry. The given key should
693 /// not already be in the map, and the given ScalarParts should be
694 /// correctly sized for the current unroll and vectorization factors.
695 const ScalarParts &initScalar(Value *Key, const ScalarParts &Entry) {
696 assert(!hasScalar(Key) && "Scalar entry already initialized")((!hasScalar(Key) && "Scalar entry already initialized"
) ? static_cast<void> (0) : __assert_fail ("!hasScalar(Key) && \"Scalar entry already initialized\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 696, __PRETTY_FUNCTION__))
;
697 assert(Entry.size() == UF &&((Entry.size() == UF && all_of(make_range(Entry.begin
(), Entry.end()), [&](const SmallVectorImpl<Value *>
&Values) -> bool { return Values.size() == VF; }) &&
"ScalarParts has wrong dimensions") ? static_cast<void>
(0) : __assert_fail ("Entry.size() == UF && all_of(make_range(Entry.begin(), Entry.end()), [&](const SmallVectorImpl<Value *> &Values) -> bool { return Values.size() == VF; }) && \"ScalarParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 702, __PRETTY_FUNCTION__))
698 all_of(make_range(Entry.begin(), Entry.end()),((Entry.size() == UF && all_of(make_range(Entry.begin
(), Entry.end()), [&](const SmallVectorImpl<Value *>
&Values) -> bool { return Values.size() == VF; }) &&
"ScalarParts has wrong dimensions") ? static_cast<void>
(0) : __assert_fail ("Entry.size() == UF && all_of(make_range(Entry.begin(), Entry.end()), [&](const SmallVectorImpl<Value *> &Values) -> bool { return Values.size() == VF; }) && \"ScalarParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 702, __PRETTY_FUNCTION__))
699 [&](const SmallVectorImpl<Value *> &Values) -> bool {((Entry.size() == UF && all_of(make_range(Entry.begin
(), Entry.end()), [&](const SmallVectorImpl<Value *>
&Values) -> bool { return Values.size() == VF; }) &&
"ScalarParts has wrong dimensions") ? static_cast<void>
(0) : __assert_fail ("Entry.size() == UF && all_of(make_range(Entry.begin(), Entry.end()), [&](const SmallVectorImpl<Value *> &Values) -> bool { return Values.size() == VF; }) && \"ScalarParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 702, __PRETTY_FUNCTION__))
700 return Values.size() == VF;((Entry.size() == UF && all_of(make_range(Entry.begin
(), Entry.end()), [&](const SmallVectorImpl<Value *>
&Values) -> bool { return Values.size() == VF; }) &&
"ScalarParts has wrong dimensions") ? static_cast<void>
(0) : __assert_fail ("Entry.size() == UF && all_of(make_range(Entry.begin(), Entry.end()), [&](const SmallVectorImpl<Value *> &Values) -> bool { return Values.size() == VF; }) && \"ScalarParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 702, __PRETTY_FUNCTION__))
701 }) &&((Entry.size() == UF && all_of(make_range(Entry.begin
(), Entry.end()), [&](const SmallVectorImpl<Value *>
&Values) -> bool { return Values.size() == VF; }) &&
"ScalarParts has wrong dimensions") ? static_cast<void>
(0) : __assert_fail ("Entry.size() == UF && all_of(make_range(Entry.begin(), Entry.end()), [&](const SmallVectorImpl<Value *> &Values) -> bool { return Values.size() == VF; }) && \"ScalarParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 702, __PRETTY_FUNCTION__))
702 "ScalarParts has wrong dimensions")((Entry.size() == UF && all_of(make_range(Entry.begin
(), Entry.end()), [&](const SmallVectorImpl<Value *>
&Values) -> bool { return Values.size() == VF; }) &&
"ScalarParts has wrong dimensions") ? static_cast<void>
(0) : __assert_fail ("Entry.size() == UF && all_of(make_range(Entry.begin(), Entry.end()), [&](const SmallVectorImpl<Value *> &Values) -> bool { return Values.size() == VF; }) && \"ScalarParts has wrong dimensions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 702, __PRETTY_FUNCTION__))
;
703 ScalarMapStorage[Key] = Entry;
704 return ScalarMapStorage[Key];
705 }
706
707 /// \return A reference to the vector map entry corresponding to \p Key.
708 /// The key should already be in the map. This function should only be used
709 /// when it's necessary to update values that have already been vectorized.
710 /// This is the case for "fix-up" operations including type truncation and
711 /// the second phase of recurrence vectorization. If a non-const reference
712 /// isn't required, getVectorValue should be used instead.
713 VectorParts &getVector(Value *Key) {
714 assert(hasVector(Key) && "Vector entry not initialized")((hasVector(Key) && "Vector entry not initialized") ?
static_cast<void> (0) : __assert_fail ("hasVector(Key) && \"Vector entry not initialized\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 714, __PRETTY_FUNCTION__))
;
715 return VectorMapStorage.find(Key)->second;
716 }
717
718 /// Retrieve an entry from the vector or scalar maps. The preferred way to
719 /// access an existing mapped entry is with getVectorValue or
720 /// getScalarValue from InnerLoopVectorizer. Until those functions can be
721 /// moved inside ValueMap, we have to declare them as friends.
722 friend const VectorParts &InnerLoopVectorizer::getVectorValue(Value *V);
723 friend Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
724 unsigned Lane);
725
726 private:
727 /// The unroll factor. Each entry in the vector map contains UF vector
728 /// values.
729 unsigned UF;
730
731 /// The vectorization factor. Each entry in the scalar map contains UF x VF
732 /// scalar values.
733 unsigned VF;
734
735 /// The vector and scalar map storage. We use std::map and not DenseMap
736 /// because insertions to DenseMap invalidate its iterators.
737 std::map<Value *, VectorParts> VectorMapStorage;
738 std::map<Value *, ScalarParts> ScalarMapStorage;
739 };
740
741 /// The original loop.
742 Loop *OrigLoop;
743 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
744 /// dynamic knowledge to simplify SCEV expressions and converts them to a
745 /// more usable form.
746 PredicatedScalarEvolution &PSE;
747 /// Loop Info.
748 LoopInfo *LI;
749 /// Dominator Tree.
750 DominatorTree *DT;
751 /// Alias Analysis.
752 AliasAnalysis *AA;
753 /// Target Library Info.
754 const TargetLibraryInfo *TLI;
755 /// Target Transform Info.
756 const TargetTransformInfo *TTI;
757 /// Assumption Cache.
758 AssumptionCache *AC;
759 /// Interface to emit optimization remarks.
760 OptimizationRemarkEmitter *ORE;
761
762 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
763 /// used.
764 ///
765 /// This is currently only used to add no-alias metadata based on the
766 /// memchecks. The actually versioning is performed manually.
767 std::unique_ptr<LoopVersioning> LVer;
768
769 /// The vectorization SIMD factor to use. Each vector will have this many
770 /// vector elements.
771 unsigned VF;
772
773protected:
774 /// The vectorization unroll factor to use. Each scalar is vectorized to this
775 /// many different vector instructions.
776 unsigned UF;
777
778 /// The builder that we use
779 IRBuilder<> Builder;
780
781 // --- Vectorization state ---
782
783 /// The vector-loop preheader.
784 BasicBlock *LoopVectorPreHeader;
785 /// The scalar-loop preheader.
786 BasicBlock *LoopScalarPreHeader;
787 /// Middle Block between the vector and the scalar.
788 BasicBlock *LoopMiddleBlock;
789 /// The ExitBlock of the scalar loop.
790 BasicBlock *LoopExitBlock;
791 /// The vector loop body.
792 BasicBlock *LoopVectorBody;
793 /// The scalar loop body.
794 BasicBlock *LoopScalarBody;
795 /// A list of all bypass blocks. The first block is the entry of the loop.
796 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
797
798 /// The new Induction variable which was added to the new block.
799 PHINode *Induction;
800 /// The induction variable of the old basic block.
801 PHINode *OldInduction;
802
803 /// Maps values from the original loop to their corresponding values in the
804 /// vectorized loop. A key value can map to either vector values, scalar
805 /// values or both kinds of values, depending on whether the key was
806 /// vectorized and scalarized.
807 ValueMap VectorLoopValueMap;
808
809 /// Store instructions that should be predicated, as a pair
810 /// <StoreInst, Predicate>
811 SmallVector<std::pair<Instruction *, Value *>, 4> PredicatedInstructions;
812 EdgeMaskCache MaskCache;
813 /// Trip count of the original loop.
814 Value *TripCount;
815 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
816 Value *VectorTripCount;
817
818 /// The legality analysis.
819 LoopVectorizationLegality *Legal;
820
821 /// The profitablity analysis.
822 LoopVectorizationCostModel *Cost;
823
824 // Record whether runtime checks are added.
825 bool AddedSafetyChecks;
826
827 // Holds instructions from the original loop whose counterparts in the
828 // vectorized loop would be trivially dead if generated. For example,
829 // original induction update instructions can become dead because we
830 // separately emit induction "steps" when generating code for the new loop.
831 // Similarly, we create a new latch condition when setting up the structure
832 // of the new loop, so the old one can become dead.
833 SmallPtrSet<Instruction *, 4> DeadInstructions;
834
835 // Holds the end values for each induction variable. We save the end values
836 // so we can later fix-up the external users of the induction variables.
837 DenseMap<PHINode *, Value *> IVEndValues;
838};
839
840class InnerLoopUnroller : public InnerLoopVectorizer {
841public:
842 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
843 LoopInfo *LI, DominatorTree *DT,
844 const TargetLibraryInfo *TLI,
845 const TargetTransformInfo *TTI, AssumptionCache *AC,
846 OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
847 LoopVectorizationLegality *LVL,
848 LoopVectorizationCostModel *CM)
849 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
850 UnrollFactor, LVL, CM) {}
851
852private:
853 void vectorizeMemoryInstruction(Instruction *Instr) override;
854 Value *getBroadcastInstrs(Value *V) override;
855 Value *getStepVector(Value *Val, int StartIdx, Value *Step,
856 Instruction::BinaryOps Opcode =
857 Instruction::BinaryOpsEnd) override;
858 Value *reverseVector(Value *Vec) override;
859};
860
861/// \brief Look for a meaningful debug location on the instruction or it's
862/// operands.
863static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
864 if (!I)
865 return I;
866
867 DebugLoc Empty;
868 if (I->getDebugLoc() != Empty)
869 return I;
870
871 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
872 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
873 if (OpInst->getDebugLoc() != Empty)
874 return OpInst;
875 }
876
877 return I;
878}
879
880void InnerLoopVectorizer::setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
881 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) {
882 const DILocation *DIL = Inst->getDebugLoc();
883 if (DIL && Inst->getFunction()->isDebugInfoForProfiling())
884 B.SetCurrentDebugLocation(DIL->cloneWithDuplicationFactor(UF * VF));
885 else
886 B.SetCurrentDebugLocation(DIL);
887 } else
888 B.SetCurrentDebugLocation(DebugLoc());
889}
890
891#ifndef NDEBUG
892/// \return string containing a file name and a line # for the given loop.
893static std::string getDebugLocString(const Loop *L) {
894 std::string Result;
895 if (L) {
896 raw_string_ostream OS(Result);
897 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
898 LoopDbgLoc.print(OS);
899 else
900 // Just print the module name.
901 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
902 OS.flush();
903 }
904 return Result;
905}
906#endif
907
908void InnerLoopVectorizer::addNewMetadata(Instruction *To,
909 const Instruction *Orig) {
910 // If the loop was versioned with memchecks, add the corresponding no-alias
911 // metadata.
912 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
913 LVer->annotateInstWithNoAlias(To, Orig);
914}
915
916void InnerLoopVectorizer::addMetadata(Instruction *To,
917 Instruction *From) {
918 propagateMetadata(To, From);
919 addNewMetadata(To, From);
920}
921
922void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
923 Instruction *From) {
924 for (Value *V : To) {
925 if (Instruction *I = dyn_cast<Instruction>(V))
926 addMetadata(I, From);
927 }
928}
929
930/// \brief The group of interleaved loads/stores sharing the same stride and
931/// close to each other.
932///
933/// Each member in this group has an index starting from 0, and the largest
934/// index should be less than interleaved factor, which is equal to the absolute
935/// value of the access's stride.
936///
937/// E.g. An interleaved load group of factor 4:
938/// for (unsigned i = 0; i < 1024; i+=4) {
939/// a = A[i]; // Member of index 0
940/// b = A[i+1]; // Member of index 1
941/// d = A[i+3]; // Member of index 3
942/// ...
943/// }
944///
945/// An interleaved store group of factor 4:
946/// for (unsigned i = 0; i < 1024; i+=4) {
947/// ...
948/// A[i] = a; // Member of index 0
949/// A[i+1] = b; // Member of index 1
950/// A[i+2] = c; // Member of index 2
951/// A[i+3] = d; // Member of index 3
952/// }
953///
954/// Note: the interleaved load group could have gaps (missing members), but
955/// the interleaved store group doesn't allow gaps.
956class InterleaveGroup {
957public:
958 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
959 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
960 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 960, __PRETTY_FUNCTION__))
;
961
962 Factor = std::abs(Stride);
963 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 963, __PRETTY_FUNCTION__))
;
964
965 Reverse = Stride < 0;
966 Members[0] = Instr;
967 }
968
969 bool isReverse() const { return Reverse; }
970 unsigned getFactor() const { return Factor; }
971 unsigned getAlignment() const { return Align; }
972 unsigned getNumMembers() const { return Members.size(); }
973
974 /// \brief Try to insert a new member \p Instr with index \p Index and
975 /// alignment \p NewAlign. The index is related to the leader and it could be
976 /// negative if it is the new leader.
977 ///
978 /// \returns false if the instruction doesn't belong to the group.
979 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
980 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 980, __PRETTY_FUNCTION__))
;
981
982 int Key = Index + SmallestKey;
983
984 // Skip if there is already a member with the same index.
985 if (Members.count(Key))
986 return false;
987
988 if (Key > LargestKey) {
989 // The largest index is always less than the interleave factor.
990 if (Index >= static_cast<int>(Factor))
991 return false;
992
993 LargestKey = Key;
994 } else if (Key < SmallestKey) {
995 // The largest index is always less than the interleave factor.
996 if (LargestKey - Key >= static_cast<int>(Factor))
997 return false;
998
999 SmallestKey = Key;
1000 }
1001
1002 // It's always safe to select the minimum alignment.
1003 Align = std::min(Align, NewAlign);
1004 Members[Key] = Instr;
1005 return true;
1006 }
1007
1008 /// \brief Get the member with the given index \p Index
1009 ///
1010 /// \returns nullptr if contains no such member.
1011 Instruction *getMember(unsigned Index) const {
1012 int Key = SmallestKey + Index;
1013 if (!Members.count(Key))
1014 return nullptr;
1015
1016 return Members.find(Key)->second;
1017 }
1018
1019 /// \brief Get the index for the given member. Unlike the key in the member
1020 /// map, the index starts from 0.
1021 unsigned getIndex(Instruction *Instr) const {
1022 for (auto I : Members)
1023 if (I.second == Instr)
1024 return I.first - SmallestKey;
1025
1026 llvm_unreachable("InterleaveGroup contains no such member")::llvm::llvm_unreachable_internal("InterleaveGroup contains no such member"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1026)
;
1027 }
1028
1029 Instruction *getInsertPos() const { return InsertPos; }
1030 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
1031
1032private:
1033 unsigned Factor; // Interleave Factor.
1034 bool Reverse;
1035 unsigned Align;
1036 DenseMap<int, Instruction *> Members;
1037 int SmallestKey;
1038 int LargestKey;
1039
1040 // To avoid breaking dependences, vectorized instructions of an interleave
1041 // group should be inserted at either the first load or the last store in
1042 // program order.
1043 //
1044 // E.g. %even = load i32 // Insert Position
1045 // %add = add i32 %even // Use of %even
1046 // %odd = load i32
1047 //
1048 // store i32 %even
1049 // %odd = add i32 // Def of %odd
1050 // store i32 %odd // Insert Position
1051 Instruction *InsertPos;
1052};
1053
1054/// \brief Drive the analysis of interleaved memory accesses in the loop.
1055///
1056/// Use this class to analyze interleaved accesses only when we can vectorize
1057/// a loop. Otherwise it's meaningless to do analysis as the vectorization
1058/// on interleaved accesses is unsafe.
1059///
1060/// The analysis collects interleave groups and records the relationships
1061/// between the member and the group in a map.
1062class InterleavedAccessInfo {
1063public:
1064 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1065 DominatorTree *DT, LoopInfo *LI)
1066 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
1067 RequiresScalarEpilogue(false) {}
1068
1069 ~InterleavedAccessInfo() {
1070 SmallSet<InterleaveGroup *, 4> DelSet;
1071 // Avoid releasing a pointer twice.
1072 for (auto &I : InterleaveGroupMap)
1073 DelSet.insert(I.second);
1074 for (auto *Ptr : DelSet)
1075 delete Ptr;
1076 }
1077
1078 /// \brief Analyze the interleaved accesses and collect them in interleave
1079 /// groups. Substitute symbolic strides using \p Strides.
1080 void analyzeInterleaving(const ValueToValueMap &Strides);
1081
1082 /// \brief Check if \p Instr belongs to any interleave group.
1083 bool isInterleaved(Instruction *Instr) const {
1084 return InterleaveGroupMap.count(Instr);
1085 }
1086
1087 /// \brief Return the maximum interleave factor of all interleaved groups.
1088 unsigned getMaxInterleaveFactor() const {
1089 unsigned MaxFactor = 1;
1090 for (auto &Entry : InterleaveGroupMap)
1091 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
1092 return MaxFactor;
1093 }
1094
1095 /// \brief Get the interleave group that \p Instr belongs to.
1096 ///
1097 /// \returns nullptr if doesn't have such group.
1098 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1099 if (InterleaveGroupMap.count(Instr))
1100 return InterleaveGroupMap.find(Instr)->second;
1101 return nullptr;
1102 }
1103
1104 /// \brief Returns true if an interleaved group that may access memory
1105 /// out-of-bounds requires a scalar epilogue iteration for correctness.
1106 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1107
1108 /// \brief Initialize the LoopAccessInfo used for dependence checking.
1109 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1110
1111private:
1112 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1113 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1114 /// The interleaved access analysis can also add new predicates (for example
1115 /// by versioning strides of pointers).
1116 PredicatedScalarEvolution &PSE;
1117 Loop *TheLoop;
1118 DominatorTree *DT;
1119 LoopInfo *LI;
1120 const LoopAccessInfo *LAI;
1121
1122 /// True if the loop may contain non-reversed interleaved groups with
1123 /// out-of-bounds accesses. We ensure we don't speculatively access memory
1124 /// out-of-bounds by executing at least one scalar epilogue iteration.
1125 bool RequiresScalarEpilogue;
1126
1127 /// Holds the relationships between the members and the interleave group.
1128 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1129
1130 /// Holds dependences among the memory accesses in the loop. It maps a source
1131 /// access to a set of dependent sink accesses.
1132 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
1133
1134 /// \brief The descriptor for a strided memory access.
1135 struct StrideDescriptor {
1136 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1137 unsigned Align)
1138 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1139
1140 StrideDescriptor() = default;
1141
1142 // The access's stride. It is negative for a reverse access.
1143 int64_t Stride = 0;
1144 const SCEV *Scev = nullptr; // The scalar expression of this access
1145 uint64_t Size = 0; // The size of the memory object.
1146 unsigned Align = 0; // The alignment of this access.
1147 };
1148
1149 /// \brief A type for holding instructions and their stride descriptors.
1150 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
1151
1152 /// \brief Create a new interleave group with the given instruction \p Instr,
1153 /// stride \p Stride and alignment \p Align.
1154 ///
1155 /// \returns the newly created interleave group.
1156 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1157 unsigned Align) {
1158 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1159, __PRETTY_FUNCTION__))
1159 "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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1159, __PRETTY_FUNCTION__))
;
1160 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1161 return InterleaveGroupMap[Instr];
1162 }
1163
1164 /// \brief Release the group and remove all the relationships.
1165 void releaseGroup(InterleaveGroup *Group) {
1166 for (unsigned i = 0; i < Group->getFactor(); i++)
1167 if (Instruction *Member = Group->getMember(i))
1168 InterleaveGroupMap.erase(Member);
1169
1170 delete Group;
1171 }
1172
1173 /// \brief Collect all the accesses with a constant stride in program order.
1174 void collectConstStrideAccesses(
1175 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1176 const ValueToValueMap &Strides);
1177
1178 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1179 static bool isStrided(int Stride) {
1180 unsigned Factor = std::abs(Stride);
1181 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1182 }
1183
1184 /// \brief Returns true if \p BB is a predicated block.
1185 bool isPredicated(BasicBlock *BB) const {
1186 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1187 }
1188
1189 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1190 bool areDependencesValid() const {
1191 return LAI && LAI->getDepChecker().getDependences();
1192 }
1193
1194 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1195 /// necessary, when constructing interleaved groups.
1196 ///
1197 /// \p A must precede \p B in program order. We return false if reordering is
1198 /// not necessary or is prevented because \p A and \p B may be dependent.
1199 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1200 StrideEntry *B) const {
1201
1202 // Code motion for interleaved accesses can potentially hoist strided loads
1203 // and sink strided stores. The code below checks the legality of the
1204 // following two conditions:
1205 //
1206 // 1. Potentially moving a strided load (B) before any store (A) that
1207 // precedes B, or
1208 //
1209 // 2. Potentially moving a strided store (A) after any load or store (B)
1210 // that A precedes.
1211 //
1212 // It's legal to reorder A and B if we know there isn't a dependence from A
1213 // to B. Note that this determination is conservative since some
1214 // dependences could potentially be reordered safely.
1215
1216 // A is potentially the source of a dependence.
1217 auto *Src = A->first;
1218 auto SrcDes = A->second;
1219
1220 // B is potentially the sink of a dependence.
1221 auto *Sink = B->first;
1222 auto SinkDes = B->second;
1223
1224 // Code motion for interleaved accesses can't violate WAR dependences.
1225 // Thus, reordering is legal if the source isn't a write.
1226 if (!Src->mayWriteToMemory())
1227 return true;
1228
1229 // At least one of the accesses must be strided.
1230 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1231 return true;
1232
1233 // If dependence information is not available from LoopAccessInfo,
1234 // conservatively assume the instructions can't be reordered.
1235 if (!areDependencesValid())
1236 return false;
1237
1238 // If we know there is a dependence from source to sink, assume the
1239 // instructions can't be reordered. Otherwise, reordering is legal.
1240 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1241 }
1242
1243 /// \brief Collect the dependences from LoopAccessInfo.
1244 ///
1245 /// We process the dependences once during the interleaved access analysis to
1246 /// enable constant-time dependence queries.
1247 void collectDependences() {
1248 if (!areDependencesValid())
1249 return;
1250 auto *Deps = LAI->getDepChecker().getDependences();
1251 for (auto Dep : *Deps)
1252 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1253 }
1254};
1255
1256/// Utility class for getting and setting loop vectorizer hints in the form
1257/// of loop metadata.
1258/// This class keeps a number of loop annotations locally (as member variables)
1259/// and can, upon request, write them back as metadata on the loop. It will
1260/// initially scan the loop for existing metadata, and will update the local
1261/// values based on information in the loop.
1262/// We cannot write all values to metadata, as the mere presence of some info,
1263/// for example 'force', means a decision has been made. So, we need to be
1264/// careful NOT to add them if the user hasn't specifically asked so.
1265class LoopVectorizeHints {
1266 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1267
1268 /// Hint - associates name and validation with the hint value.
1269 struct Hint {
1270 const char *Name;
1271 unsigned Value; // This may have to change for non-numeric values.
1272 HintKind Kind;
1273
1274 Hint(const char *Name, unsigned Value, HintKind Kind)
1275 : Name(Name), Value(Value), Kind(Kind) {}
1276
1277 bool validate(unsigned Val) {
1278 switch (Kind) {
1279 case HK_WIDTH:
1280 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1281 case HK_UNROLL:
1282 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1283 case HK_FORCE:
1284 return (Val <= 1);
1285 }
1286 return false;
1287 }
1288 };
1289
1290 /// Vectorization width.
1291 Hint Width;
1292 /// Vectorization interleave factor.
1293 Hint Interleave;
1294 /// Vectorization forced
1295 Hint Force;
1296
1297 /// Return the loop metadata prefix.
1298 static StringRef Prefix() { return "llvm.loop."; }
1299
1300 /// True if there is any unsafe math in the loop.
1301 bool PotentiallyUnsafe;
1302
1303public:
1304 enum ForceKind {
1305 FK_Undefined = -1, ///< Not selected.
1306 FK_Disabled = 0, ///< Forcing disabled.
1307 FK_Enabled = 1, ///< Forcing enabled.
1308 };
1309
1310 LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1311 OptimizationRemarkEmitter &ORE)
1312 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1313 HK_WIDTH),
1314 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1315 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1316 PotentiallyUnsafe(false), TheLoop(L), ORE(ORE) {
1317 // Populate values with existing loop metadata.
1318 getHintsFromMetadata();
1319
1320 // force-vector-interleave overrides DisableInterleaving.
1321 if (VectorizerParams::isInterleaveForced())
1322 Interleave.Value = VectorizerParams::VectorizationInterleave;
1323
1324 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 (false)
1325 << "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 (false)
;
1326 }
1327
1328 /// Mark the loop L as already vectorized by setting the width to 1.
1329 void setAlreadyVectorized() {
1330 Width.Value = Interleave.Value = 1;
1331 Hint Hints[] = {Width, Interleave};
1332 writeHintsToMetadata(Hints);
1333 }
1334
1335 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1336 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1337 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 (false)
;
1338 emitRemarkWithHints();
1339 return false;
1340 }
1341
1342 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1343 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 (false)
;
1344 emitRemarkWithHints();
1345 return false;
1346 }
1347
1348 if (getWidth() == 1 && getInterleave() == 1) {
1349 // FIXME: Add a separate metadata to indicate when the loop has already
1350 // been vectorized instead of setting width and count to 1.
1351 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 (false)
;
1352 // FIXME: Add interleave.disable metadata. This will allow
1353 // vectorize.disable to be used without disabling the pass and errors
1354 // to differentiate between disabled vectorization and a width of 1.
1355 ORE.emit(OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1356 "AllDisabled", L->getStartLoc(),
1357 L->getHeader())
1358 << "loop not vectorized: vectorization and interleaving are "
1359 "explicitly disabled, or vectorize width and interleave "
1360 "count are both set to 1");
1361 return false;
1362 }
1363
1364 return true;
1365 }
1366
1367 /// Dumps all the hint information.
1368 void emitRemarkWithHints() const {
1369 using namespace ore;
1370 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1371 ORE.emit(OptimizationRemarkMissed(LV_NAME"loop-vectorize", "MissedExplicitlyDisabled",
1372 TheLoop->getStartLoc(),
1373 TheLoop->getHeader())
1374 << "loop not vectorized: vectorization is explicitly disabled");
1375 else {
1376 OptimizationRemarkMissed R(LV_NAME"loop-vectorize", "MissedDetails",
1377 TheLoop->getStartLoc(), TheLoop->getHeader());
1378 R << "loop not vectorized";
1379 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1380 R << " (Force=" << NV("Force", true);
1381 if (Width.Value != 0)
1382 R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1383 if (Interleave.Value != 0)
1384 R << ", Interleave Count=" << NV("InterleaveCount", Interleave.Value);
1385 R << ")";
1386 }
1387 ORE.emit(R);
1388 }
1389 }
1390
1391 unsigned getWidth() const { return Width.Value; }
1392 unsigned getInterleave() const { return Interleave.Value; }
1393 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1394
1395 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1396 /// pass name to force the frontend to print the diagnostic.
1397 const char *vectorizeAnalysisPassName() const {
1398 if (getWidth() == 1)
1399 return LV_NAME"loop-vectorize";
1400 if (getForce() == LoopVectorizeHints::FK_Disabled)
1401 return LV_NAME"loop-vectorize";
1402 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1403 return LV_NAME"loop-vectorize";
1404 return OptimizationRemarkAnalysis::AlwaysPrint;
1405 }
1406
1407 bool allowReordering() const {
1408 // When enabling loop hints are provided we allow the vectorizer to change
1409 // the order of operations that is given by the scalar loop. This is not
1410 // enabled by default because can be unsafe or inefficient. For example,
1411 // reordering floating-point operations will change the way round-off
1412 // error accumulates in the loop.
1413 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1414 }
1415
1416 bool isPotentiallyUnsafe() const {
1417 // Avoid FP vectorization if the target is unsure about proper support.
1418 // This may be related to the SIMD unit in the target not handling
1419 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1420 // Otherwise, a sequence of vectorized loops, even without reduction,
1421 // could lead to different end results on the destination vectors.
1422 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1423 }
1424
1425 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1426
1427private:
1428 /// Find hints specified in the loop metadata and update local values.
1429 void getHintsFromMetadata() {
1430 MDNode *LoopID = TheLoop->getLoopID();
1431 if (!LoopID)
1432 return;
1433
1434 // First operand should refer to the loop id itself.
1435 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1435, __PRETTY_FUNCTION__))
;
1436 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1436, __PRETTY_FUNCTION__))
;
1437
1438 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1439 const MDString *S = nullptr;
1440 SmallVector<Metadata *, 4> Args;
1441
1442 // The expected hint is either a MDString or a MDNode with the first
1443 // operand a MDString.
1444 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1445 if (!MD || MD->getNumOperands() == 0)
1446 continue;
1447 S = dyn_cast<MDString>(MD->getOperand(0));
1448 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1449 Args.push_back(MD->getOperand(i));
1450 } else {
1451 S = dyn_cast<MDString>(LoopID->getOperand(i));
1452 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1452, __PRETTY_FUNCTION__))
;
1453 }
1454
1455 if (!S)
1456 continue;
1457
1458 // Check if the hint starts with the loop metadata prefix.
1459 StringRef Name = S->getString();
1460 if (Args.size() == 1)
1461 setHint(Name, Args[0]);
1462 }
1463 }
1464
1465 /// Checks string hint with one operand and set value if valid.
1466 void setHint(StringRef Name, Metadata *Arg) {
1467 if (!Name.startswith(Prefix()))
1468 return;
1469 Name = Name.substr(Prefix().size(), StringRef::npos);
1470
1471 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1472 if (!C)
1473 return;
1474 unsigned Val = C->getZExtValue();
1475
1476 Hint *Hints[] = {&Width, &Interleave, &Force};
1477 for (auto H : Hints) {
1478 if (Name == H->Name) {
1479 if (H->validate(Val))
1480 H->Value = Val;
1481 else
1482 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: ignoring invalid hint '"
<< Name << "'\n"; } } while (false)
;
1483 break;
1484 }
1485 }
1486 }
1487
1488 /// Create a new hint from name / value pair.
1489 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1490 LLVMContext &Context = TheLoop->getHeader()->getContext();
1491 Metadata *MDs[] = {MDString::get(Context, Name),
1492 ConstantAsMetadata::get(
1493 ConstantInt::get(Type::getInt32Ty(Context), V))};
1494 return MDNode::get(Context, MDs);
1495 }
1496
1497 /// Matches metadata with hint name.
1498 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1499 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1500 if (!Name)
1501 return false;
1502
1503 for (auto H : HintTypes)
1504 if (Name->getString().endswith(H.Name))
1505 return true;
1506 return false;
1507 }
1508
1509 /// Sets current hints into loop metadata, keeping other values intact.
1510 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1511 if (HintTypes.size() == 0)
1512 return;
1513
1514 // Reserve the first element to LoopID (see below).
1515 SmallVector<Metadata *, 4> MDs(1);
1516 // If the loop already has metadata, then ignore the existing operands.
1517 MDNode *LoopID = TheLoop->getLoopID();
1518 if (LoopID) {
1519 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1520 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1521 // If node in update list, ignore old value.
1522 if (!matchesHintMetadataName(Node, HintTypes))
1523 MDs.push_back(Node);
1524 }
1525 }
1526
1527 // Now, add the missing hints.
1528 for (auto H : HintTypes)
1529 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1530
1531 // Replace current metadata node with new one.
1532 LLVMContext &Context = TheLoop->getHeader()->getContext();
1533 MDNode *NewLoopID = MDNode::get(Context, MDs);
1534 // Set operand 0 to refer to the loop id itself.
1535 NewLoopID->replaceOperandWith(0, NewLoopID);
1536
1537 TheLoop->setLoopID(NewLoopID);
1538 }
1539
1540 /// The loop these hints belong to.
1541 const Loop *TheLoop;
1542
1543 /// Interface to emit optimization remarks.
1544 OptimizationRemarkEmitter &ORE;
1545};
1546
1547static void emitMissedWarning(Function *F, Loop *L,
1548 const LoopVectorizeHints &LH,
1549 OptimizationRemarkEmitter *ORE) {
1550 LH.emitRemarkWithHints();
1551
1552 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1553 if (LH.getWidth() != 1)
1554 ORE->emit(DiagnosticInfoOptimizationFailure(
1555 DEBUG_TYPE"loop-vectorize", "FailedRequestedVectorization",
1556 L->getStartLoc(), L->getHeader())
1557 << "loop not vectorized: "
1558 << "failed explicitly specified loop vectorization");
1559 else if (LH.getInterleave() != 1)
1560 ORE->emit(DiagnosticInfoOptimizationFailure(
1561 DEBUG_TYPE"loop-vectorize", "FailedRequestedInterleaving", L->getStartLoc(),
1562 L->getHeader())
1563 << "loop not interleaved: "
1564 << "failed explicitly specified loop interleaving");
1565 }
1566}
1567
1568/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1569/// to what vectorization factor.
1570/// This class does not look at the profitability of vectorization, only the
1571/// legality. This class has two main kinds of checks:
1572/// * Memory checks - The code in canVectorizeMemory checks if vectorization
1573/// will change the order of memory accesses in a way that will change the
1574/// correctness of the program.
1575/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1576/// checks for a number of different conditions, such as the availability of a
1577/// single induction variable, that all types are supported and vectorize-able,
1578/// etc. This code reflects the capabilities of InnerLoopVectorizer.
1579/// This class is also used by InnerLoopVectorizer for identifying
1580/// induction variable and the different reduction variables.
1581class LoopVectorizationLegality {
1582public:
1583 LoopVectorizationLegality(
1584 Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
1585 TargetLibraryInfo *TLI, AliasAnalysis *AA, Function *F,
1586 const TargetTransformInfo *TTI,
1587 std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1588 OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R,
1589 LoopVectorizeHints *H)
1590 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TTI(TTI), DT(DT),
1591 GetLAA(GetLAA), LAI(nullptr), ORE(ORE), InterleaveInfo(PSE, L, DT, LI),
1592 PrimaryInduction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1593 Requirements(R), Hints(H) {}
1594
1595 /// ReductionList contains the reduction descriptors for all
1596 /// of the reductions that were found in the loop.
1597 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1598
1599 /// InductionList saves induction variables and maps them to the
1600 /// induction descriptor.
1601 typedef MapVector<PHINode *, InductionDescriptor> InductionList;
1602
1603 /// RecurrenceSet contains the phi nodes that are recurrences other than
1604 /// inductions and reductions.
1605 typedef SmallPtrSet<const PHINode *, 8> RecurrenceSet;
1606
1607 /// Returns true if it is legal to vectorize this loop.
1608 /// This does not mean that it is profitable to vectorize this
1609 /// loop, only that it is legal to do so.
1610 bool canVectorize();
1611
1612 /// Returns the primary induction variable.
1613 PHINode *getPrimaryInduction() { return PrimaryInduction; }
1614
1615 /// Returns the reduction variables found in the loop.
1616 ReductionList *getReductionVars() { return &Reductions; }
1617
1618 /// Returns the induction variables found in the loop.
1619 InductionList *getInductionVars() { return &Inductions; }
1620
1621 /// Return the first-order recurrences found in the loop.
1622 RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1623
1624 /// Returns the widest induction type.
1625 Type *getWidestInductionType() { return WidestIndTy; }
1626
1627 /// Returns True if V is an induction variable in this loop.
1628 bool isInductionVariable(const Value *V);
1629
1630 /// Returns True if PN is a reduction variable in this loop.
1631 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1632
1633 /// Returns True if Phi is a first-order recurrence in this loop.
1634 bool isFirstOrderRecurrence(const PHINode *Phi);
1635
1636 /// Return true if the block BB needs to be predicated in order for the loop
1637 /// to be vectorized.
1638 bool blockNeedsPredication(BasicBlock *BB);
1639
1640 /// Check if this pointer is consecutive when vectorizing. This happens
1641 /// when the last index of the GEP is the induction variable, or that the
1642 /// pointer itself is an induction variable.
1643 /// This check allows us to vectorize A[idx] into a wide load/store.
1644 /// Returns:
1645 /// 0 - Stride is unknown or non-consecutive.
1646 /// 1 - Address is consecutive.
1647 /// -1 - Address is consecutive, and decreasing.
1648 int isConsecutivePtr(Value *Ptr);
1649
1650 /// Returns true if the value V is uniform within the loop.
1651 bool isUniform(Value *V);
1652
1653 /// Returns the information that we collected about runtime memory check.
1654 const RuntimePointerChecking *getRuntimePointerChecking() const {
1655 return LAI->getRuntimePointerChecking();
1656 }
1657
1658 const LoopAccessInfo *getLAI() const { return LAI; }
1659
1660 /// \brief Check if \p Instr belongs to any interleaved access group.
1661 bool isAccessInterleaved(Instruction *Instr) {
1662 return InterleaveInfo.isInterleaved(Instr);
1663 }
1664
1665 /// \brief Return the maximum interleave factor of all interleaved groups.
1666 unsigned getMaxInterleaveFactor() const {
1667 return InterleaveInfo.getMaxInterleaveFactor();
1668 }
1669
1670 /// \brief Get the interleaved access group that \p Instr belongs to.
1671 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1672 return InterleaveInfo.getInterleaveGroup(Instr);
1673 }
1674
1675 /// \brief Returns true if an interleaved group requires a scalar iteration
1676 /// to handle accesses with gaps.
1677 bool requiresScalarEpilogue() const {
1678 return InterleaveInfo.requiresScalarEpilogue();
1679 }
1680
1681 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1682
1683 bool hasStride(Value *V) { return LAI->hasStride(V); }
1684
1685 /// Returns true if the target machine supports masked store operation
1686 /// for the given \p DataType and kind of access to \p Ptr.
1687 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1688 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1689 }
1690 /// Returns true if the target machine supports masked load operation
1691 /// for the given \p DataType and kind of access to \p Ptr.
1692 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1693 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1694 }
1695 /// Returns true if the target machine supports masked scatter operation
1696 /// for the given \p DataType.
1697 bool isLegalMaskedScatter(Type *DataType) {
1698 return TTI->isLegalMaskedScatter(DataType);
1699 }
1700 /// Returns true if the target machine supports masked gather operation
1701 /// for the given \p DataType.
1702 bool isLegalMaskedGather(Type *DataType) {
1703 return TTI->isLegalMaskedGather(DataType);
1704 }
1705 /// Returns true if the target machine can represent \p V as a masked gather
1706 /// or scatter operation.
1707 bool isLegalGatherOrScatter(Value *V) {
1708 auto *LI = dyn_cast<LoadInst>(V);
1709 auto *SI = dyn_cast<StoreInst>(V);
1710 if (!LI && !SI)
1711 return false;
1712 auto *Ptr = getPointerOperand(V);
1713 auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1714 return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1715 }
1716
1717 /// Returns true if vector representation of the instruction \p I
1718 /// requires mask.
1719 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1720 unsigned getNumStores() const { return LAI->getNumStores(); }
1721 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1722 unsigned getNumPredStores() const { return NumPredStores; }
1723
1724 /// Returns true if \p I is an instruction that will be scalarized with
1725 /// predication. Such instructions include conditional stores and
1726 /// instructions that may divide by zero.
1727 bool isScalarWithPredication(Instruction *I);
1728
1729 /// Returns true if \p I is a memory instruction with consecutive memory
1730 /// access that can be widened.
1731 bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
1732
1733private:
1734 /// Check if a single basic block loop is vectorizable.
1735 /// At this point we know that this is a loop with a constant trip count
1736 /// and we only need to check individual instructions.
1737 bool canVectorizeInstrs();
1738
1739 /// When we vectorize loops we may change the order in which
1740 /// we read and write from memory. This method checks if it is
1741 /// legal to vectorize the code, considering only memory constrains.
1742 /// Returns true if the loop is vectorizable
1743 bool canVectorizeMemory();
1744
1745 /// Return true if we can vectorize this loop using the IF-conversion
1746 /// transformation.
1747 bool canVectorizeWithIfConvert();
1748
1749 /// Return true if all of the instructions in the block can be speculatively
1750 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1751 /// and we know that we can read from them without segfault.
1752 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1753
1754 /// Updates the vectorization state by adding \p Phi to the inductions list.
1755 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1756 /// better choice for the main induction than the existing one.
1757 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1758 SmallPtrSetImpl<Value *> &AllowedExit);
1759
1760 /// Create an analysis remark that explains why vectorization failed
1761 ///
1762 /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1763 /// an instruction that prevents vectorization. Otherwise the loop is used
1764 /// for the location of the remark. \return the remark object that can be
1765 /// streamed to.
1766 OptimizationRemarkAnalysis
1767 createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1768 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1769 RemarkName, TheLoop, I);
1770 }
1771
1772 /// \brief If an access has a symbolic strides, this maps the pointer value to
1773 /// the stride symbol.
1774 const ValueToValueMap *getSymbolicStrides() {
1775 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1776 // it's collected. This happens from canVectorizeWithIfConvert, when the
1777 // pointer is checked to reference consecutive elements suitable for a
1778 // masked access.
1779 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1780 }
1781
1782 unsigned NumPredStores;
1783
1784 /// The loop that we evaluate.
1785 Loop *TheLoop;
1786 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1787 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1788 /// of existing SCEV assumptions. The analysis will also add a minimal set
1789 /// of new predicates if this is required to enable vectorization and
1790 /// unrolling.
1791 PredicatedScalarEvolution &PSE;
1792 /// Target Library Info.
1793 TargetLibraryInfo *TLI;
1794 /// Target Transform Info
1795 const TargetTransformInfo *TTI;
1796 /// Dominator Tree.
1797 DominatorTree *DT;
1798 // LoopAccess analysis.
1799 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1800 // And the loop-accesses info corresponding to this loop. This pointer is
1801 // null until canVectorizeMemory sets it up.
1802 const LoopAccessInfo *LAI;
1803 /// Interface to emit optimization remarks.
1804 OptimizationRemarkEmitter *ORE;
1805
1806 /// The interleave access information contains groups of interleaved accesses
1807 /// with the same stride and close to each other.
1808 InterleavedAccessInfo InterleaveInfo;
1809
1810 // --- vectorization state --- //
1811
1812 /// Holds the primary induction variable. This is the counter of the
1813 /// loop.
1814 PHINode *PrimaryInduction;
1815 /// Holds the reduction variables.
1816 ReductionList Reductions;
1817 /// Holds all of the induction variables that we found in the loop.
1818 /// Notice that inductions don't need to start at zero and that induction
1819 /// variables can be pointers.
1820 InductionList Inductions;
1821 /// Holds the phi nodes that are first-order recurrences.
1822 RecurrenceSet FirstOrderRecurrences;
1823 /// Holds the widest induction type encountered.
1824 Type *WidestIndTy;
1825
1826 /// Allowed outside users. This holds the induction and reduction
1827 /// vars which can be accessed from outside the loop.
1828 SmallPtrSet<Value *, 4> AllowedExit;
1829
1830 /// Can we assume the absence of NaNs.
1831 bool HasFunNoNaNAttr;
1832
1833 /// Vectorization requirements that will go through late-evaluation.
1834 LoopVectorizationRequirements *Requirements;
1835
1836 /// Used to emit an analysis of any legality issues.
1837 LoopVectorizeHints *Hints;
1838
1839 /// While vectorizing these instructions we have to generate a
1840 /// call to the appropriate masked intrinsic
1841 SmallPtrSet<const Instruction *, 8> MaskedOp;
1842};
1843
1844/// LoopVectorizationCostModel - estimates the expected speedups due to
1845/// vectorization.
1846/// In many cases vectorization is not profitable. This can happen because of
1847/// a number of reasons. In this class we mainly attempt to predict the
1848/// expected speedup/slowdowns due to the supported instruction set. We use the
1849/// TargetTransformInfo to query the different backends for the cost of
1850/// different operations.
1851class LoopVectorizationCostModel {
1852public:
1853 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1854 LoopInfo *LI, LoopVectorizationLegality *Legal,
1855 const TargetTransformInfo &TTI,
1856 const TargetLibraryInfo *TLI, DemandedBits *DB,
1857 AssumptionCache *AC,
1858 OptimizationRemarkEmitter *ORE, const Function *F,
1859 const LoopVectorizeHints *Hints)
1860 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1861 AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1862
1863 /// \return An upper bound for the vectorization factor, or None if
1864 /// vectorization should be avoided up front.
1865 Optional<unsigned> computeMaxVF(bool OptForSize);
1866
1867 /// Information about vectorization costs
1868 struct VectorizationFactor {
1869 unsigned Width; // Vector width with best cost
1870 unsigned Cost; // Cost of the loop with that width
1871 };
1872 /// \return The most profitable vectorization factor and the cost of that VF.
1873 /// This method checks every power of two up to MaxVF. If UserVF is not ZERO
1874 /// then this vectorization factor will be selected if vectorization is
1875 /// possible.
1876 VectorizationFactor selectVectorizationFactor(unsigned MaxVF);
1877
1878 /// Setup cost-based decisions for user vectorization factor.
1879 void selectUserVectorizationFactor(unsigned UserVF) {
1880 collectUniformsAndScalars(UserVF);
1881 collectInstsToScalarize(UserVF);
1882 }
1883
1884 /// \return The size (in bits) of the smallest and widest types in the code
1885 /// that needs to be vectorized. We ignore values that remain scalar such as
1886 /// 64 bit loop indices.
1887 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1888
1889 /// \return The desired interleave count.
1890 /// If interleave count has been specified by metadata it will be returned.
1891 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1892 /// are the selected vectorization factor and the cost of the selected VF.
1893 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1894 unsigned LoopCost);
1895
1896 /// Memory access instruction may be vectorized in more than one way.
1897 /// Form of instruction after vectorization depends on cost.
1898 /// This function takes cost-based decisions for Load/Store instructions
1899 /// and collects them in a map. This decisions map is used for building
1900 /// the lists of loop-uniform and loop-scalar instructions.
1901 /// The calculated cost is saved with widening decision in order to
1902 /// avoid redundant calculations.
1903 void setCostBasedWideningDecision(unsigned VF);
1904
1905 /// \brief A struct that represents some properties of the register usage
1906 /// of a loop.
1907 struct RegisterUsage {
1908 /// Holds the number of loop invariant values that are used in the loop.
1909 unsigned LoopInvariantRegs;
1910 /// Holds the maximum number of concurrent live intervals in the loop.
1911 unsigned MaxLocalUsers;
1912 /// Holds the number of instructions in the loop.
1913 unsigned NumInstructions;
1914 };
1915
1916 /// \return Returns information about the register usages of the loop for the
1917 /// given vectorization factors.
1918 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1919
1920 /// Collect values we want to ignore in the cost model.
1921 void collectValuesToIgnore();
1922
1923 /// \returns The smallest bitwidth each instruction can be represented with.
1924 /// The vector equivalents of these instructions should be truncated to this
1925 /// type.
1926 const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1927 return MinBWs;
1928 }
1929
1930 /// \returns True if it is more profitable to scalarize instruction \p I for
1931 /// vectorization factor \p VF.
1932 bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1933 auto Scalars = InstsToScalarize.find(VF);
1934 assert(Scalars != InstsToScalarize.end() &&((Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability"
) ? static_cast<void> (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1935, __PRETTY_FUNCTION__))
1935 "VF not yet analyzed for scalarization profitability")((Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability"
) ? static_cast<void> (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1935, __PRETTY_FUNCTION__))
;
1936 return Scalars->second.count(I);
1937 }
1938
1939 /// Returns true if \p I is known to be uniform after vectorization.
1940 bool isUniformAfterVectorization(Instruction *I, unsigned VF) const {
1941 if (VF == 1)
1942 return true;
1943 assert(Uniforms.count(VF) && "VF not yet analyzed for uniformity")((Uniforms.count(VF) && "VF not yet analyzed for uniformity"
) ? static_cast<void> (0) : __assert_fail ("Uniforms.count(VF) && \"VF not yet analyzed for uniformity\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1943, __PRETTY_FUNCTION__))
;
1944 auto UniformsPerVF = Uniforms.find(VF);
1945 return UniformsPerVF->second.count(I);
1946 }
1947
1948 /// Returns true if \p I is known to be scalar after vectorization.
1949 bool isScalarAfterVectorization(Instruction *I, unsigned VF) const {
1950 if (VF == 1)
1951 return true;
1952 assert(Scalars.count(VF) && "Scalar values are not calculated for VF")((Scalars.count(VF) && "Scalar values are not calculated for VF"
) ? static_cast<void> (0) : __assert_fail ("Scalars.count(VF) && \"Scalar values are not calculated for VF\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1952, __PRETTY_FUNCTION__))
;
1953 auto ScalarsPerVF = Scalars.find(VF);
1954 return ScalarsPerVF->second.count(I);
1955 }
1956
1957 /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1958 /// for vectorization factor \p VF.
1959 bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1960 return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1961 !isScalarAfterVectorization(I, VF);
1962 }
1963
1964 /// Decision that was taken during cost calculation for memory instruction.
1965 enum InstWidening {
1966 CM_Unknown,
1967 CM_Widen,
1968 CM_Interleave,
1969 CM_GatherScatter,
1970 CM_Scalarize
1971 };
1972
1973 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1974 /// instruction \p I and vector width \p VF.
1975 void setWideningDecision(Instruction *I, unsigned VF, InstWidening W,
1976 unsigned Cost) {
1977 assert(VF >= 2 && "Expected VF >=2")((VF >= 2 && "Expected VF >=2") ? static_cast<
void> (0) : __assert_fail ("VF >= 2 && \"Expected VF >=2\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1977, __PRETTY_FUNCTION__))
;
1978 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1979 }
1980
1981 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1982 /// interleaving group \p Grp and vector width \p VF.
1983 void setWideningDecision(const InterleaveGroup *Grp, unsigned VF,
1984 InstWidening W, unsigned Cost) {
1985 assert(VF >= 2 && "Expected VF >=2")((VF >= 2 && "Expected VF >=2") ? static_cast<
void> (0) : __assert_fail ("VF >= 2 && \"Expected VF >=2\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1985, __PRETTY_FUNCTION__))
;
1986 /// Broadcast this decicion to all instructions inside the group.
1987 /// But the cost will be assigned to one instruction only.
1988 for (unsigned i = 0; i < Grp->getFactor(); ++i) {
1989 if (auto *I = Grp->getMember(i)) {
1990 if (Grp->getInsertPos() == I)
1991 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1992 else
1993 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0);
1994 }
1995 }
1996 }
1997
1998 /// Return the cost model decision for the given instruction \p I and vector
1999 /// width \p VF. Return CM_Unknown if this instruction did not pass
2000 /// through the cost modeling.
2001 InstWidening getWideningDecision(Instruction *I, unsigned VF) {
2002 assert(VF >= 2 && "Expected VF >=2")((VF >= 2 && "Expected VF >=2") ? static_cast<
void> (0) : __assert_fail ("VF >= 2 && \"Expected VF >=2\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2002, __PRETTY_FUNCTION__))
;
2003 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
2004 auto Itr = WideningDecisions.find(InstOnVF);
2005 if (Itr == WideningDecisions.end())
2006 return CM_Unknown;
2007 return Itr->second.first;
2008 }
2009
2010 /// Return the vectorization cost for the given instruction \p I and vector
2011 /// width \p VF.
2012 unsigned getWideningCost(Instruction *I, unsigned VF) {
2013 assert(VF >= 2 && "Expected VF >=2")((VF >= 2 && "Expected VF >=2") ? static_cast<
void> (0) : __assert_fail ("VF >= 2 && \"Expected VF >=2\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2013, __PRETTY_FUNCTION__))
;
2014 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
2015 assert(WideningDecisions.count(InstOnVF) && "The cost is not calculated")((WideningDecisions.count(InstOnVF) && "The cost is not calculated"
) ? static_cast<void> (0) : __assert_fail ("WideningDecisions.count(InstOnVF) && \"The cost is not calculated\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2015, __PRETTY_FUNCTION__))
;
2016 return WideningDecisions[InstOnVF].second;
2017 }
2018
2019 /// Return True if instruction \p I is an optimizable truncate whose operand
2020 /// is an induction variable. Such a truncate will be removed by adding a new
2021 /// induction variable with the destination type.
2022 bool isOptimizableIVTruncate(Instruction *I, unsigned VF) {
2023
2024 // If the instruction is not a truncate, return false.
2025 auto *Trunc = dyn_cast<TruncInst>(I);
2026 if (!Trunc)
2027 return false;
2028
2029 // Get the source and destination types of the truncate.
2030 Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF);
2031 Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF);
2032
2033 // If the truncate is free for the given types, return false. Replacing a
2034 // free truncate with an induction variable would add an induction variable
2035 // update instruction to each iteration of the loop. We exclude from this
2036 // check the primary induction variable since it will need an update
2037 // instruction regardless.
2038 Value *Op = Trunc->getOperand(0);
2039 if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))
2040 return false;
2041
2042 // If the truncated value is not an induction variable, return false.
2043 return Legal->isInductionVariable(Op);
2044 }
2045
2046private:
2047 /// \return An upper bound for the vectorization factor, larger than zero.
2048 /// One is returned if vectorization should best be avoided due to cost.
2049 unsigned computeFeasibleMaxVF(bool OptForSize);
2050
2051 /// The vectorization cost is a combination of the cost itself and a boolean
2052 /// indicating whether any of the contributing operations will actually
2053 /// operate on
2054 /// vector values after type legalization in the backend. If this latter value
2055 /// is
2056 /// false, then all operations will be scalarized (i.e. no vectorization has
2057 /// actually taken place).
2058 typedef std::pair<unsigned, bool> VectorizationCostTy;
2059
2060 /// Returns the expected execution cost. The unit of the cost does
2061 /// not matter because we use the 'cost' units to compare different
2062 /// vector widths. The cost that is returned is *not* normalized by
2063 /// the factor width.
2064 VectorizationCostTy expectedCost(unsigned VF);
2065
2066 /// Returns the execution time cost of an instruction for a given vector
2067 /// width. Vector width of one means scalar.
2068 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
2069
2070 /// The cost-computation logic from getInstructionCost which provides
2071 /// the vector type as an output parameter.
2072 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
2073
2074 /// Calculate vectorization cost of memory instruction \p I.
2075 unsigned getMemoryInstructionCost(Instruction *I, unsigned VF);
2076
2077 /// The cost computation for scalarized memory instruction.
2078 unsigned getMemInstScalarizationCost(Instruction *I, unsigned VF);
2079
2080 /// The cost computation for interleaving group of memory instructions.
2081 unsigned getInterleaveGroupCost(Instruction *I, unsigned VF);
2082
2083 /// The cost computation for Gather/Scatter instruction.
2084 unsigned getGatherScatterCost(Instruction *I, unsigned VF);
2085
2086 /// The cost computation for widening instruction \p I with consecutive
2087 /// memory access.
2088 unsigned getConsecutiveMemOpCost(Instruction *I, unsigned VF);
2089
2090 /// The cost calculation for Load instruction \p I with uniform pointer -
2091 /// scalar load + broadcast.
2092 unsigned getUniformMemOpCost(Instruction *I, unsigned VF);
2093
2094 /// Returns whether the instruction is a load or store and will be a emitted
2095 /// as a vector operation.
2096 bool isConsecutiveLoadOrStore(Instruction *I);
2097
2098 /// Create an analysis remark that explains why vectorization failed
2099 ///
2100 /// \p RemarkName is the identifier for the remark. \return the remark object
2101 /// that can be streamed to.
2102 OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
2103 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
2104 RemarkName, TheLoop);
2105 }
2106
2107 /// Map of scalar integer values to the smallest bitwidth they can be legally
2108 /// represented as. The vector equivalents of these values should be truncated
2109 /// to this type.
2110 MapVector<Instruction *, uint64_t> MinBWs;
2111
2112 /// A type representing the costs for instructions if they were to be
2113 /// scalarized rather than vectorized. The entries are Instruction-Cost
2114 /// pairs.
2115 typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
2116
2117 /// A map holding scalar costs for different vectorization factors. The
2118 /// presence of a cost for an instruction in the mapping indicates that the
2119 /// instruction will be scalarized when vectorizing with the associated
2120 /// vectorization factor. The entries are VF-ScalarCostTy pairs.
2121 DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
2122
2123 /// Holds the instructions known to be uniform after vectorization.
2124 /// The data is collected per VF.
2125 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Uniforms;
2126
2127 /// Holds the instructions known to be scalar after vectorization.
2128 /// The data is collected per VF.
2129 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Scalars;
2130
2131 /// Returns the expected difference in cost from scalarizing the expression
2132 /// feeding a predicated instruction \p PredInst. The instructions to
2133 /// scalarize and their scalar costs are collected in \p ScalarCosts. A
2134 /// non-negative return value implies the expression will be scalarized.
2135 /// Currently, only single-use chains are considered for scalarization.
2136 int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
2137 unsigned VF);
2138
2139 /// Collects the instructions to scalarize for each predicated instruction in
2140 /// the loop.
2141 void collectInstsToScalarize(unsigned VF);
2142
2143 /// Collect the instructions that are uniform after vectorization. An
2144 /// instruction is uniform if we represent it with a single scalar value in
2145 /// the vectorized loop corresponding to each vector iteration. Examples of
2146 /// uniform instructions include pointer operands of consecutive or
2147 /// interleaved memory accesses. Note that although uniformity implies an
2148 /// instruction will be scalar, the reverse is not true. In general, a
2149 /// scalarized instruction will be represented by VF scalar values in the
2150 /// vectorized loop, each corresponding to an iteration of the original
2151 /// scalar loop.
2152 void collectLoopUniforms(unsigned VF);
2153
2154 /// Collect the instructions that are scalar after vectorization. An
2155 /// instruction is scalar if it is known to be uniform or will be scalarized
2156 /// during vectorization. Non-uniform scalarized instructions will be
2157 /// represented by VF values in the vectorized loop, each corresponding to an
2158 /// iteration of the original scalar loop.
2159 void collectLoopScalars(unsigned VF);
2160
2161 /// Collect Uniform and Scalar values for the given \p VF.
2162 /// The sets depend on CM decision for Load/Store instructions
2163 /// that may be vectorized as interleave, gather-scatter or scalarized.
2164 void collectUniformsAndScalars(unsigned VF) {
2165 // Do the analysis once.
2166 if (VF == 1 || Uniforms.count(VF))
2167 return;
2168 setCostBasedWideningDecision(VF);
2169 collectLoopUniforms(VF);
2170 collectLoopScalars(VF);
2171 }
2172
2173 /// Keeps cost model vectorization decision and cost for instructions.
2174 /// Right now it is used for memory instructions only.
2175 typedef DenseMap<std::pair<Instruction *, unsigned>,
2176 std::pair<InstWidening, unsigned>>
2177 DecisionList;
2178
2179 DecisionList WideningDecisions;
2180
2181public:
2182 /// The loop that we evaluate.
2183 Loop *TheLoop;
2184 /// Predicated scalar evolution analysis.
2185 PredicatedScalarEvolution &PSE;
2186 /// Loop Info analysis.
2187 LoopInfo *LI;
2188 /// Vectorization legality.
2189 LoopVectorizationLegality *Legal;
2190 /// Vector target information.
2191 const TargetTransformInfo &TTI;
2192 /// Target Library Info.
2193 const TargetLibraryInfo *TLI;
2194 /// Demanded bits analysis.
2195 DemandedBits *DB;
2196 /// Assumption cache.
2197 AssumptionCache *AC;
2198 /// Interface to emit optimization remarks.
2199 OptimizationRemarkEmitter *ORE;
2200
2201 const Function *TheFunction;
2202 /// Loop Vectorize Hint.
2203 const LoopVectorizeHints *Hints;
2204 /// Values to ignore in the cost model.
2205 SmallPtrSet<const Value *, 16> ValuesToIgnore;
2206 /// Values to ignore in the cost model when VF > 1.
2207 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2208};
2209
2210/// LoopVectorizationPlanner - drives the vectorization process after having
2211/// passed Legality checks.
2212class LoopVectorizationPlanner {
2213public:
2214 LoopVectorizationPlanner(LoopVectorizationCostModel &CM) : CM(CM) {}
2215
2216 ~LoopVectorizationPlanner() {}
2217
2218 /// Plan how to best vectorize, return the best VF and its cost.
2219 LoopVectorizationCostModel::VectorizationFactor plan(bool OptForSize,
2220 unsigned UserVF);
2221
2222private:
2223 /// The profitablity analysis.
2224 LoopVectorizationCostModel &CM;
2225};
2226
2227/// \brief This holds vectorization requirements that must be verified late in
2228/// the process. The requirements are set by legalize and costmodel. Once
2229/// vectorization has been determined to be possible and profitable the
2230/// requirements can be verified by looking for metadata or compiler options.
2231/// For example, some loops require FP commutativity which is only allowed if
2232/// vectorization is explicitly specified or if the fast-math compiler option
2233/// has been provided.
2234/// Late evaluation of these requirements allows helpful diagnostics to be
2235/// composed that tells the user what need to be done to vectorize the loop. For
2236/// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2237/// evaluation should be used only when diagnostics can generated that can be
2238/// followed by a non-expert user.
2239class LoopVectorizationRequirements {
2240public:
2241 LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE)
2242 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr), ORE(ORE) {}
2243
2244 void addUnsafeAlgebraInst(Instruction *I) {
2245 // First unsafe algebra instruction.
2246 if (!UnsafeAlgebraInst)
2247 UnsafeAlgebraInst = I;
2248 }
2249
2250 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2251
2252 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2253 const char *PassName = Hints.vectorizeAnalysisPassName();
2254 bool Failed = false;
2255 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2256 ORE.emit(
2257 OptimizationRemarkAnalysisFPCommute(PassName, "CantReorderFPOps",
2258 UnsafeAlgebraInst->getDebugLoc(),
2259 UnsafeAlgebraInst->getParent())
2260 << "loop not vectorized: cannot prove it is safe to reorder "
2261 "floating-point operations");
2262 Failed = true;
2263 }
2264
2265 // Test if runtime memcheck thresholds are exceeded.
2266 bool PragmaThresholdReached =
2267 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2268 bool ThresholdReached =
2269 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2270 if ((ThresholdReached && !Hints.allowReordering()) ||
2271 PragmaThresholdReached) {
2272 ORE.emit(OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2273 L->getStartLoc(),
2274 L->getHeader())
2275 << "loop not vectorized: cannot prove it is safe to reorder "
2276 "memory operations");
2277 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 (false)
;
2278 Failed = true;
2279 }
2280
2281 return Failed;
2282 }
2283
2284private:
2285 unsigned NumRuntimePointerChecks;
2286 Instruction *UnsafeAlgebraInst;
2287
2288 /// Interface to emit optimization remarks.
2289 OptimizationRemarkEmitter &ORE;
2290};
2291
2292static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
2293 if (L.empty()) {
2294 if (!hasCyclesInLoopBody(L))
2295 V.push_back(&L);
2296 return;
2297 }
2298 for (Loop *InnerL : L)
2299 addAcyclicInnerLoop(*InnerL, V);
2300}
2301
2302/// The LoopVectorize Pass.
2303struct LoopVectorize : public FunctionPass {
2304 /// Pass identification, replacement for typeid
2305 static char ID;
2306
2307 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2308 : FunctionPass(ID) {
2309 Impl.DisableUnrolling = NoUnrolling;
2310 Impl.AlwaysVectorize = AlwaysVectorize;
2311 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
2312 }
2313
2314 LoopVectorizePass Impl;
2315
2316 bool runOnFunction(Function &F) override {
2317 if (skipFunction(F))
2318 return false;
2319
2320 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2321 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2322 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2323 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2324 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2325 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2326 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2327 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2328 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2329 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2330 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2331 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2332
2333 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2334 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2335
2336 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2337 GetLAA, *ORE);
2338 }
2339
2340 void getAnalysisUsage(AnalysisUsage &AU) const override {
2341 AU.addRequired<AssumptionCacheTracker>();
2342 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2343 AU.addRequired<DominatorTreeWrapperPass>();
2344 AU.addRequired<LoopInfoWrapperPass>();
2345 AU.addRequired<ScalarEvolutionWrapperPass>();
2346 AU.addRequired<TargetTransformInfoWrapperPass>();
2347 AU.addRequired<AAResultsWrapperPass>();
2348 AU.addRequired<LoopAccessLegacyAnalysis>();
2349 AU.addRequired<DemandedBitsWrapperPass>();
2350 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2351 AU.addPreserved<LoopInfoWrapperPass>();
2352 AU.addPreserved<DominatorTreeWrapperPass>();
2353 AU.addPreserved<BasicAAWrapperPass>();
2354 AU.addPreserved<GlobalsAAWrapperPass>();
2355 }
2356};
2357
2358} // end anonymous namespace
2359
2360//===----------------------------------------------------------------------===//
2361// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2362// LoopVectorizationCostModel and LoopVectorizationPlanner.
2363//===----------------------------------------------------------------------===//
2364
2365Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
2366 // We need to place the broadcast of invariant variables outside the loop.
2367 Instruction *Instr = dyn_cast<Instruction>(V);
2368 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2369 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2370
2371 // Place the code for broadcasting invariant variables in the new preheader.
2372 IRBuilder<>::InsertPointGuard Guard(Builder);
2373 if (Invariant)
2374 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2375
2376 // Broadcast the scalar into all locations in the vector.
2377 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2378
2379 return Shuf;
2380}
2381
2382void InnerLoopVectorizer::createVectorIntOrFpInductionPHI(
2383 const InductionDescriptor &II, Value *Step, Instruction *EntryVal) {
2384 Value *Start = II.getStartValue();
2385
2386 // Construct the initial value of the vector IV in the vector loop preheader
2387 auto CurrIP = Builder.saveIP();
2388 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2389 if (isa<TruncInst>(EntryVal)) {
2390 assert(Start->getType()->isIntegerTy() &&((Start->getType()->isIntegerTy() && "Truncation requires an integer type"
) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2391, __PRETTY_FUNCTION__))
2391 "Truncation requires an integer type")((Start->getType()->isIntegerTy() && "Truncation requires an integer type"
) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2391, __PRETTY_FUNCTION__))
;
2392 auto *TruncType = cast<IntegerType>(EntryVal->getType());
2393 Step = Builder.CreateTrunc(Step, TruncType);
2394 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2395 }
2396 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2397 Value *SteppedStart =
2398 getStepVector(SplatStart, 0, Step, II.getInductionOpcode());
2399
2400 // We create vector phi nodes for both integer and floating-point induction
2401 // variables. Here, we determine the kind of arithmetic we will perform.
2402 Instruction::BinaryOps AddOp;
2403 Instruction::BinaryOps MulOp;
2404 if (Step->getType()->isIntegerTy()) {
2405 AddOp = Instruction::Add;
2406 MulOp = Instruction::Mul;
2407 } else {
2408 AddOp = II.getInductionOpcode();
2409 MulOp = Instruction::FMul;
2410 }
2411
2412 // Multiply the vectorization factor by the step using integer or
2413 // floating-point arithmetic as appropriate.
2414 Value *ConstVF = getSignedIntOrFpConstant(Step->getType(), VF);
2415 Value *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, Step, ConstVF));
2416
2417 // Create a vector splat to use in the induction update.
2418 //
2419 // FIXME: If the step is non-constant, we create the vector splat with
2420 // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
2421 // handle a constant vector splat.
2422 Value *SplatVF = isa<Constant>(Mul)
2423 ? ConstantVector::getSplat(VF, cast<Constant>(Mul))
2424 : Builder.CreateVectorSplat(VF, Mul);
2425 Builder.restoreIP(CurrIP);
2426
2427 // We may need to add the step a number of times, depending on the unroll
2428 // factor. The last of those goes into the PHI.
2429 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2430 &*LoopVectorBody->getFirstInsertionPt());
2431 Instruction *LastInduction = VecInd;
2432 VectorParts Entry(UF);
2433 for (unsigned Part = 0; Part < UF; ++Part) {
2434 Entry[Part] = LastInduction;
2435 LastInduction = cast<Instruction>(addFastMathFlag(
2436 Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")));
2437 }
2438 VectorLoopValueMap.initVector(EntryVal, Entry);
2439 if (isa<TruncInst>(EntryVal))
2440 addMetadata(Entry, EntryVal);
2441
2442 // Move the last step to the end of the latch block. This ensures consistent
2443 // placement of all induction updates.
2444 auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2445 auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2446 auto *ICmp = cast<Instruction>(Br->getCondition());
2447 LastInduction->moveBefore(ICmp);
2448 LastInduction->setName("vec.ind.next");
2449
2450 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2451 VecInd->addIncoming(LastInduction, LoopVectorLatch);
2452}
2453
2454bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
2455 return Cost->isScalarAfterVectorization(I, VF) ||
2456 Cost->isProfitableToScalarize(I, VF);
2457}
2458
2459bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
2460 if (shouldScalarizeInstruction(IV))
2461 return true;
2462 auto isScalarInst = [&](User *U) -> bool {
2463 auto *I = cast<Instruction>(U);
2464 return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2465 };
2466 return any_of(IV->users(), isScalarInst);
2467}
2468
2469void InnerLoopVectorizer::widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc) {
2470
2471 assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&(((IV->getType()->isIntegerTy() || IV != OldInduction) &&
"Primary induction variable must have an integer type") ? static_cast
<void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2472, __PRETTY_FUNCTION__))
2472 "Primary induction variable must have an integer type")(((IV->getType()->isIntegerTy() || IV != OldInduction) &&
"Primary induction variable must have an integer type") ? static_cast
<void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2472, __PRETTY_FUNCTION__))
;
2473
2474 auto II = Legal->getInductionVars()->find(IV);
2475 assert(II != Legal->getInductionVars()->end() && "IV is not an induction")((II != Legal->getInductionVars()->end() && "IV is not an induction"
) ? static_cast<void> (0) : __assert_fail ("II != Legal->getInductionVars()->end() && \"IV is not an induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2475, __PRETTY_FUNCTION__))
;
2476
2477 auto ID = II->second;
2478 assert(IV->getType() == ID.getStartValue()->getType() && "Types must match")((IV->getType() == ID.getStartValue()->getType() &&
"Types must match") ? static_cast<void> (0) : __assert_fail
("IV->getType() == ID.getStartValue()->getType() && \"Types must match\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2478, __PRETTY_FUNCTION__))
;
2479
2480 // The scalar value to broadcast. This will be derived from the canonical
2481 // induction variable.
2482 Value *ScalarIV = nullptr;
2483
2484 // The value from the original loop to which we are mapping the new induction
2485 // variable.
2486 Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2487
2488 // True if we have vectorized the induction variable.
2489 auto VectorizedIV = false;
2490
2491 // Determine if we want a scalar version of the induction variable. This is
2492 // true if the induction variable itself is not widened, or if it has at
2493 // least one user in the loop that is not widened.
2494 auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2495
2496 // Generate code for the induction step. Note that induction steps are
2497 // required to be loop-invariant
2498 assert(PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) &&((PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) &&
"Induction step should be loop invariant") ? static_cast<
void> (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) && \"Induction step should be loop invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2499, __PRETTY_FUNCTION__))
2499 "Induction step should be loop invariant")((PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) &&
"Induction step should be loop invariant") ? static_cast<
void> (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) && \"Induction step should be loop invariant\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2499, __PRETTY_FUNCTION__))
;
2500 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2501 Value *Step = nullptr;
2502 if (PSE.getSE()->isSCEVable(IV->getType())) {
2503 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2504 Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2505 LoopVectorPreHeader->getTerminator());
2506 } else {
2507 Step = cast<SCEVUnknown>(ID.getStep())->getValue();
2508 }
2509
2510 // Try to create a new independent vector induction variable. If we can't
2511 // create the phi node, we will splat the scalar induction variable in each
2512 // loop iteration.
2513 if (VF > 1 && !shouldScalarizeInstruction(EntryVal)) {
2514 createVectorIntOrFpInductionPHI(ID, Step, EntryVal);
2515 VectorizedIV = true;
2516 }
2517
2518 // If we haven't yet vectorized the induction variable, or if we will create
2519 // a scalar one, we need to define the scalar induction variable and step
2520 // values. If we were given a truncation type, truncate the canonical
2521 // induction variable and step. Otherwise, derive these values from the
2522 // induction descriptor.
2523 if (!VectorizedIV || NeedsScalarIV) {
2524 if (Trunc) {
2525 auto *TruncType = cast<IntegerType>(Trunc->getType());
2526 assert(Step->getType()->isIntegerTy() &&((Step->getType()->isIntegerTy() && "Truncation requires an integer step"
) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2527, __PRETTY_FUNCTION__))
2527 "Truncation requires an integer step")((Step->getType()->isIntegerTy() && "Truncation requires an integer step"
) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2527, __PRETTY_FUNCTION__))
;
2528 ScalarIV = Builder.CreateCast(Instruction::Trunc, Induction, TruncType);
2529 Step = Builder.CreateTrunc(Step, TruncType);
2530 } else {
2531 ScalarIV = Induction;
2532 if (IV != OldInduction) {
2533 ScalarIV = IV->getType()->isIntegerTy()
2534 ? Builder.CreateSExtOrTrunc(ScalarIV, IV->getType())
2535 : Builder.CreateCast(Instruction::SIToFP, Induction,
2536 IV->getType());
2537 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2538 ScalarIV->setName("offset.idx");
2539 }
2540 }
2541 }
2542
2543 // If we haven't yet vectorized the induction variable, splat the scalar
2544 // induction variable, and build the necessary step vectors.
2545 if (!VectorizedIV) {
2546 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2547 VectorParts Entry(UF);
2548 for (unsigned Part = 0; Part < UF; ++Part)
2549 Entry[Part] =
2550 getStepVector(Broadcasted, VF * Part, Step, ID.getInductionOpcode());
2551 VectorLoopValueMap.initVector(EntryVal, Entry);
2552 if (Trunc)
2553 addMetadata(Entry, Trunc);
2554 }
2555
2556 // If an induction variable is only used for counting loop iterations or
2557 // calculating addresses, it doesn't need to be widened. Create scalar steps
2558 // that can be used by instructions we will later scalarize. Note that the
2559 // addition of the scalar steps will not increase the number of instructions
2560 // in the loop in the common case prior to InstCombine. We will be trading
2561 // one vector extract for each scalar step.
2562 if (NeedsScalarIV)
2563 buildScalarSteps(ScalarIV, Step, EntryVal, ID);
2564}
2565
2566Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
2567 Instruction::BinaryOps BinOp) {
2568 // Create and check the types.
2569 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2569, __PRETTY_FUNCTION__))
;
2570 int VLen = Val->getType()->getVectorNumElements();
2571
2572 Type *STy = Val->getType()->getScalarType();
2573 assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&(((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
"Induction Step must be an integer or FP") ? static_cast<
void> (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2574, __PRETTY_FUNCTION__))
2574 "Induction Step must be an integer or FP")(((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
"Induction Step must be an integer or FP") ? static_cast<
void> (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2574, __PRETTY_FUNCTION__))
;
2575 assert(Step->getType() == STy && "Step has wrong type")((Step->getType() == STy && "Step has wrong type")
? static_cast<void> (0) : __assert_fail ("Step->getType() == STy && \"Step has wrong type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2575, __PRETTY_FUNCTION__))
;
2576
2577 SmallVector<Constant *, 8> Indices;
2578
2579 if (STy->isIntegerTy()) {
2580 // Create a vector of consecutive numbers from zero to VF.
2581 for (int i = 0; i < VLen; ++i)
2582 Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2583
2584 // Add the consecutive indices to the vector value.
2585 Constant *Cv = ConstantVector::get(Indices);
2586 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2586, __PRETTY_FUNCTION__))
;
2587 Step = Builder.CreateVectorSplat(VLen, Step);
2588 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2588, __PRETTY_FUNCTION__))
;
2589 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2590 // which can be found from the original scalar operations.
2591 Step = Builder.CreateMul(Cv, Step);
2592 return Builder.CreateAdd(Val, Step, "induction");
2593 }
2594
2595 // Floating point induction.
2596 assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&(((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
"Binary Opcode should be specified for FP induction") ? static_cast
<void> (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2597, __PRETTY_FUNCTION__))
2597 "Binary Opcode should be specified for FP induction")(((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
"Binary Opcode should be specified for FP induction") ? static_cast
<void> (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2597, __PRETTY_FUNCTION__))
;
2598 // Create a vector of consecutive numbers from zero to VF.
2599 for (int i = 0; i < VLen; ++i)
2600 Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2601
2602 // Add the consecutive indices to the vector value.
2603 Constant *Cv = ConstantVector::get(Indices);
2604
2605 Step = Builder.CreateVectorSplat(VLen, Step);
2606
2607 // Floating point operations had to be 'fast' to enable the induction.
2608 FastMathFlags Flags;
2609 Flags.setUnsafeAlgebra();
2610
2611 Value *MulOp = Builder.CreateFMul(Cv, Step);
2612 if (isa<Instruction>(MulOp))
2613 // Have to check, MulOp may be a constant
2614 cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2615
2616 Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2617 if (isa<Instruction>(BOp))
2618 cast<Instruction>(BOp)->setFastMathFlags(Flags);
2619 return BOp;
2620}
2621
2622void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2623 Value *EntryVal,
2624 const InductionDescriptor &ID) {
2625
2626 // We shouldn't have to build scalar steps if we aren't vectorizing.
2627 assert(VF > 1 && "VF should be greater than one")((VF > 1 && "VF should be greater than one") ? static_cast
<void> (0) : __assert_fail ("VF > 1 && \"VF should be greater than one\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2627, __PRETTY_FUNCTION__))
;
2628
2629 // Get the value type and ensure it and the step have the same integer type.
2630 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2631 assert(ScalarIVTy == Step->getType() &&((ScalarIVTy == Step->getType() && "Val and Step should have the same type"
) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2632, __PRETTY_FUNCTION__))
2632 "Val and Step should have the same type")((ScalarIVTy == Step->getType() && "Val and Step should have the same type"
) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2632, __PRETTY_FUNCTION__))
;
2633
2634 // We build scalar steps for both integer and floating-point induction
2635 // variables. Here, we determine the kind of arithmetic we will perform.
2636 Instruction::BinaryOps AddOp;
2637 Instruction::BinaryOps MulOp;
2638 if (ScalarIVTy->isIntegerTy()) {
2639 AddOp = Instruction::Add;
2640 MulOp = Instruction::Mul;
2641 } else {
2642 AddOp = ID.getInductionOpcode();
2643 MulOp = Instruction::FMul;
2644 }
2645
2646 // Determine the number of scalars we need to generate for each unroll
2647 // iteration. If EntryVal is uniform, we only need to generate the first
2648 // lane. Otherwise, we generate all VF values.
2649 unsigned Lanes =
2650 Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) ? 1 : VF;
2651
2652 // Compute the scalar steps and save the results in VectorLoopValueMap.
2653 ScalarParts Entry(UF);
2654 for (unsigned Part = 0; Part < UF; ++Part) {
2655 Entry[Part].resize(VF);
2656 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2657 auto *StartIdx = getSignedIntOrFpConstant(ScalarIVTy, VF * Part + Lane);
2658 auto *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, StartIdx, Step));
2659 auto *Add = addFastMathFlag(Builder.CreateBinOp(AddOp, ScalarIV, Mul));
2660 Entry[Part][Lane] = Add;
2661 }
2662 }
2663 VectorLoopValueMap.initScalar(EntryVal, Entry);
2664}
2665
2666int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2667
2668 const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2669 ValueToValueMap();
2670
2671 int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2672 if (Stride == 1 || Stride == -1)
2673 return Stride;
2674 return 0;
2675}
2676
2677bool LoopVectorizationLegality::isUniform(Value *V) {
2678 return LAI->isUniform(V);
2679}
2680
2681const InnerLoopVectorizer::VectorParts &
2682InnerLoopVectorizer::getVectorValue(Value *V) {
2683 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2683, __PRETTY_FUNCTION__))
;
2684 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2684, __PRETTY_FUNCTION__))
;
2685 assert(!V->getType()->isVoidTy() && "Type does not produce a value")((!V->getType()->isVoidTy() && "Type does not produce a value"
) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isVoidTy() && \"Type does not produce a value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2685, __PRETTY_FUNCTION__))
;
2686
2687 // If we have a stride that is replaced by one, do it here.
2688 if (Legal->hasStride(V))
2689 V = ConstantInt::get(V->getType(), 1);
2690
2691 // If we have this scalar in the map, return it.
2692 if (VectorLoopValueMap.hasVector(V))
2693 return VectorLoopValueMap.VectorMapStorage[V];
2694
2695 // If the value has not been vectorized, check if it has been scalarized
2696 // instead. If it has been scalarized, and we actually need the value in
2697 // vector form, we will construct the vector values on demand.
2698 if (VectorLoopValueMap.hasScalar(V)) {
2699
2700 // Initialize a new vector map entry.
2701 VectorParts Entry(UF);
2702
2703 // If we've scalarized a value, that value should be an instruction.
2704 auto *I = cast<Instruction>(V);
2705
2706 // If we aren't vectorizing, we can just copy the scalar map values over to
2707 // the vector map.
2708 if (VF == 1) {
2709 for (unsigned Part = 0; Part < UF; ++Part)
2710 Entry[Part] = getScalarValue(V, Part, 0);
2711 return VectorLoopValueMap.initVector(V, Entry);
2712 }
2713
2714 // Get the last scalar instruction we generated for V. If the value is
2715 // known to be uniform after vectorization, this corresponds to lane zero
2716 // of the last unroll iteration. Otherwise, the last instruction is the one
2717 // we created for the last vector lane of the last unroll iteration.
2718 unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) ? 0 : VF - 1;
2719 auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
2720
2721 // Set the insert point after the last scalarized instruction. This ensures
2722 // the insertelement sequence will directly follow the scalar definitions.
2723 auto OldIP = Builder.saveIP();
2724 auto NewIP = std::next(BasicBlock::iterator(LastInst));
2725 Builder.SetInsertPoint(&*NewIP);
2726
2727 // However, if we are vectorizing, we need to construct the vector values.
2728 // If the value is known to be uniform after vectorization, we can just
2729 // broadcast the scalar value corresponding to lane zero for each unroll
2730 // iteration. Otherwise, we construct the vector values using insertelement
2731 // instructions. Since the resulting vectors are stored in
2732 // VectorLoopValueMap, we will only generate the insertelements once.
2733 for (unsigned Part = 0; Part < UF; ++Part) {
2734 Value *VectorValue = nullptr;
2735 if (Cost->isUniformAfterVectorization(I, VF)) {
2736 VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
2737 } else {
2738 VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
2739 for (unsigned Lane = 0; Lane < VF; ++Lane)
2740 VectorValue = Builder.CreateInsertElement(
2741 VectorValue, getScalarValue(V, Part, Lane),
2742 Builder.getInt32(Lane));
2743 }
2744 Entry[Part] = VectorValue;
2745 }
2746 Builder.restoreIP(OldIP);
2747 return VectorLoopValueMap.initVector(V, Entry);
2748 }
2749
2750 // If this scalar is unknown, assume that it is a constant or that it is
2751 // loop invariant. Broadcast V and save the value for future uses.
2752 Value *B = getBroadcastInstrs(V);
2753 return VectorLoopValueMap.initVector(V, VectorParts(UF, B));
2754}
2755
2756Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
2757 unsigned Lane) {
2758
2759 // If the value is not an instruction contained in the loop, it should
2760 // already be scalar.
2761 if (OrigLoop->isLoopInvariant(V))
2762 return V;
2763
2764 assert(Lane > 0 ?((Lane > 0 ? !Cost->isUniformAfterVectorization(cast<
Instruction>(V), VF) : true && "Uniform values only have lane zero"
) ? static_cast<void> (0) : __assert_fail ("Lane > 0 ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF) : true && \"Uniform values only have lane zero\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2766, __PRETTY_FUNCTION__))
2765 !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF)((Lane > 0 ? !Cost->isUniformAfterVectorization(cast<
Instruction>(V), VF) : true && "Uniform values only have lane zero"
) ? static_cast<void> (0) : __assert_fail ("Lane > 0 ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF) : true && \"Uniform values only have lane zero\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2766, __PRETTY_FUNCTION__))
2766 : true && "Uniform values only have lane zero")((Lane > 0 ? !Cost->isUniformAfterVectorization(cast<
Instruction>(V), VF) : true && "Uniform values only have lane zero"
) ? static_cast<void> (0) : __assert_fail ("Lane > 0 ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF) : true && \"Uniform values only have lane zero\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2766, __PRETTY_FUNCTION__))
;
2767
2768 // If the value from the original loop has not been vectorized, it is
2769 // represented by UF x VF scalar values in the new loop. Return the requested
2770 // scalar value.
2771 if (VectorLoopValueMap.hasScalar(V))
2772 return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
2773
2774 // If the value has not been scalarized, get its entry in VectorLoopValueMap
2775 // for the given unroll part. If this entry is not a vector type (i.e., the
2776 // vectorization factor is one), there is no need to generate an
2777 // extractelement instruction.
2778 auto *U = getVectorValue(V)[Part];
2779 if (!U->getType()->isVectorTy()) {
2780 assert(VF == 1 && "Value not scalarized has non-vector type")((VF == 1 && "Value not scalarized has non-vector type"
) ? static_cast<void> (0) : __assert_fail ("VF == 1 && \"Value not scalarized has non-vector type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2780, __PRETTY_FUNCTION__))
;
2781 return U;
2782 }
2783
2784 // Otherwise, the value from the original loop has been vectorized and is
2785 // represented by UF vector values. Extract and return the requested scalar
2786 // value from the appropriate vector lane.
2787 return Builder.CreateExtractElement(U, Builder.getInt32(Lane));
2788}
2789
2790Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2791 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2791, __PRETTY_FUNCTION__))
;
2792 SmallVector<Constant *, 8> ShuffleMask;
2793 for (unsigned i = 0; i < VF; ++i)
2794 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2795
2796 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2797 ConstantVector::get(ShuffleMask),
2798 "reverse");
2799}
2800
2801// Try to vectorize the interleave group that \p Instr belongs to.
2802//
2803// E.g. Translate following interleaved load group (factor = 3):
2804// for (i = 0; i < N; i+=3) {
2805// R = Pic[i]; // Member of index 0
2806// G = Pic[i+1]; // Member of index 1
2807// B = Pic[i+2]; // Member of index 2
2808// ... // do something to R, G, B
2809// }
2810// To:
2811// %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2812// %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2813// %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2814// %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2815//
2816// Or translate following interleaved store group (factor = 3):
2817// for (i = 0; i < N; i+=3) {
2818// ... do something to R, G, B
2819// Pic[i] = R; // Member of index 0
2820// Pic[i+1] = G; // Member of index 1
2821// Pic[i+2] = B; // Member of index 2
2822// }
2823// To:
2824// %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2825// %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2826// %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2827// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2828// store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2829void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2830 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2831 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2831, __PRETTY_FUNCTION__))
;
2832
2833 // Skip if current instruction is not the insert position.
2834 if (Instr != Group->getInsertPos())
2835 return;
2836
2837 Value *Ptr = getPointerOperand(Instr);
2838
2839 // Prepare for the vector type of the interleaved load/store.
2840 Type *ScalarTy = getMemInstValueType(Instr);
2841 unsigned InterleaveFactor = Group->getFactor();
2842 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2843 Type *PtrTy = VecTy->getPointerTo(getMemInstAddressSpace(Instr));
2844
2845 // Prepare for the new pointers.
2846 setDebugLocFromInst(Builder, Ptr);
2847 SmallVector<Value *, 2> NewPtrs;
2848 unsigned Index = Group->getIndex(Instr);
2849
2850 // If the group is reverse, adjust the index to refer to the last vector lane
2851 // instead of the first. We adjust the index from the first vector lane,
2852 // rather than directly getting the pointer for lane VF - 1, because the
2853 // pointer operand of the interleaved access is supposed to be uniform. For
2854 // uniform instructions, we're only required to generate a value for the
2855 // first vector lane in each unroll iteration.
2856 if (Group->isReverse())
2857 Index += (VF - 1) * Group->getFactor();
2858
2859 for (unsigned Part = 0; Part < UF; Part++) {
2860 Value *NewPtr = getScalarValue(Ptr, Part, 0);
2861
2862 // Notice current instruction could be any index. Need to adjust the address
2863 // to the member of index 0.
2864 //
2865 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2866 // b = A[i]; // Member of index 0
2867 // Current pointer is pointed to A[i+1], adjust it to A[i].
2868 //
2869 // E.g. A[i+1] = a; // Member of index 1
2870 // A[i] = b; // Member of index 0
2871 // A[i+2] = c; // Member of index 2 (Current instruction)
2872 // Current pointer is pointed to A[i+2], adjust it to A[i].
2873 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2874
2875 // Cast to the vector pointer type.
2876 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2877 }
2878
2879 setDebugLocFromInst(Builder, Instr);
2880 Value *UndefVec = UndefValue::get(VecTy);
2881
2882 // Vectorize the interleaved load group.
2883 if (isa<LoadInst>(Instr)) {
2884
2885 // For each unroll part, create a wide load for the group.
2886 SmallVector<Value *, 2> NewLoads;
2887 for (unsigned Part = 0; Part < UF; Part++) {
2888 auto *NewLoad = Builder.CreateAlignedLoad(
2889 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2890 addMetadata(NewLoad, Instr);
2891 NewLoads.push_back(NewLoad);
2892 }
2893
2894 // For each member in the group, shuffle out the appropriate data from the
2895 // wide loads.
2896 for (unsigned I = 0; I < InterleaveFactor; ++I) {
2897 Instruction *Member = Group->getMember(I);
2898
2899 // Skip the gaps in the group.
2900 if (!Member)
2901 continue;
2902
2903 VectorParts Entry(UF);
2904 Constant *StrideMask = createStrideMask(Builder, I, InterleaveFactor, VF);
2905 for (unsigned Part = 0; Part < UF; Part++) {
2906 Value *StridedVec = Builder.CreateShuffleVector(
2907 NewLoads[Part], UndefVec, StrideMask, "strided.vec");
2908
2909 // If this member has different type, cast the result type.
2910 if (Member->getType() != ScalarTy) {
2911 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2912 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2913 }
2914
2915 Entry[Part] =
2916 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2917 }
2918 VectorLoopValueMap.initVector(Member, Entry);
2919 }
2920 return;
2921 }
2922
2923 // The sub vector type for current instruction.
2924 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2925
2926 // Vectorize the interleaved store group.
2927 for (unsigned Part = 0; Part < UF; Part++) {
2928 // Collect the stored vector from each member.
2929 SmallVector<Value *, 4> StoredVecs;
2930 for (unsigned i = 0; i < InterleaveFactor; i++) {
2931 // Interleaved store group doesn't allow a gap, so each index has a member
2932 Instruction *Member = Group->getMember(i);
2933 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2933, __PRETTY_FUNCTION__))
;
2934
2935 Value *StoredVec =
2936 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2937 if (Group->isReverse())
2938 StoredVec = reverseVector(StoredVec);
2939
2940 // If this member has different type, cast it to an unified type.
2941 if (StoredVec->getType() != SubVT)
2942 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2943
2944 StoredVecs.push_back(StoredVec);
2945 }
2946
2947 // Concatenate all vectors into a wide vector.
2948 Value *WideVec = concatenateVectors(Builder, StoredVecs);
2949
2950 // Interleave the elements in the wide vector.
2951 Constant *IMask = createInterleaveMask(Builder, VF, InterleaveFactor);
2952 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2953 "interleaved.vec");
2954
2955 Instruction *NewStoreInstr =
2956 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2957 addMetadata(NewStoreInstr, Instr);
2958 }
2959}
2960
2961void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2962 // Attempt to issue a wide load.
2963 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2964 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2965
2966 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2966, __PRETTY_FUNCTION__))
;
2967
2968 LoopVectorizationCostModel::InstWidening Decision =
2969 Cost->getWideningDecision(Instr, VF);
2970 assert(Decision != LoopVectorizationCostModel::CM_Unknown &&((Decision != LoopVectorizationCostModel::CM_Unknown &&
"CM decision should be taken at this point") ? static_cast<
void> (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2971, __PRETTY_FUNCTION__))
2971 "CM decision should be taken at this point")((Decision != LoopVectorizationCostModel::CM_Unknown &&
"CM decision should be taken at this point") ? static_cast<
void> (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2971, __PRETTY_FUNCTION__))
;
2972 if (Decision == LoopVectorizationCostModel::CM_Interleave)
2973 return vectorizeInterleaveGroup(Instr);
2974
2975 Type *ScalarDataTy = getMemInstValueType(Instr);
2976 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2977 Value *Ptr = getPointerOperand(Instr);
2978 unsigned Alignment = getMemInstAlignment(Instr);
2979 // An alignment of 0 means target abi alignment. We need to use the scalar's
2980 // target abi alignment in such a case.
2981 const DataLayout &DL = Instr->getModule()->getDataLayout();
2982 if (!Alignment)
2983 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2984 unsigned AddressSpace = getMemInstAddressSpace(Instr);
2985
2986 // Scalarize the memory instruction if necessary.
2987 if (Decision == LoopVectorizationCostModel::CM_Scalarize)
2988 return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
2989
2990 // Determine if the pointer operand of the access is either consecutive or
2991 // reverse consecutive.
2992 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2993 bool Reverse = ConsecutiveStride < 0;
2994 bool CreateGatherScatter =
2995 (Decision == LoopVectorizationCostModel::CM_GatherScatter);
2996
2997 VectorParts VectorGep;
2998
2999 // Handle consecutive loads/stores.
3000 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
3001 if (ConsecutiveStride) {
3002 if (Gep) {
3003 unsigned NumOperands = Gep->getNumOperands();
3004#ifndef NDEBUG
3005 // The original GEP that identified as a consecutive memory access
3006 // should have only one loop-variant operand.
3007 unsigned NumOfLoopVariantOps = 0;
3008 for (unsigned i = 0; i < NumOperands; ++i)
3009 if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)),
3010 OrigLoop))
3011 NumOfLoopVariantOps++;
3012 assert(NumOfLoopVariantOps == 1 &&((NumOfLoopVariantOps == 1 && "Consecutive GEP should have only one loop-variant operand"
) ? static_cast<void> (0) : __assert_fail ("NumOfLoopVariantOps == 1 && \"Consecutive GEP should have only one loop-variant operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3013, __PRETTY_FUNCTION__))
3013 "Consecutive GEP should have only one loop-variant operand")((NumOfLoopVariantOps == 1 && "Consecutive GEP should have only one loop-variant operand"
) ? static_cast<void> (0) : __assert_fail ("NumOfLoopVariantOps == 1 && \"Consecutive GEP should have only one loop-variant operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3013, __PRETTY_FUNCTION__))
;
3014#endif
3015 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
3016 Gep2->setName("gep.indvar");
3017
3018 // A new GEP is created for a 0-lane value of the first unroll iteration.
3019 // The GEPs for the rest of the unroll iterations are computed below as an
3020 // offset from this GEP.
3021 for (unsigned i = 0; i < NumOperands; ++i)
3022 // We can apply getScalarValue() for all GEP indices. It returns an
3023 // original value for loop-invariant operand and 0-lane for consecutive
3024 // operand.
3025 Gep2->setOperand(i, getScalarValue(Gep->getOperand(i),
3026 0, /* First unroll iteration */
3027 0 /* 0-lane of the vector */ ));
3028 setDebugLocFromInst(Builder, Gep);
3029 Ptr = Builder.Insert(Gep2);
3030
3031 } else { // No GEP
3032 setDebugLocFromInst(Builder, Ptr);
3033 Ptr = getScalarValue(Ptr, 0, 0);
3034 }
3035 } else {
3036 // At this point we should vector version of GEP for Gather or Scatter
3037 assert(CreateGatherScatter && "The instruction should be scalarized")((CreateGatherScatter && "The instruction should be scalarized"
) ? static_cast<void> (0) : __assert_fail ("CreateGatherScatter && \"The instruction should be scalarized\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3037, __PRETTY_FUNCTION__))
;
3038 if (Gep) {
3039 // Vectorizing GEP, across UF parts. We want to get a vector value for base
3040 // and each index that's defined inside the loop, even if it is
3041 // loop-invariant but wasn't hoisted out. Otherwise we want to keep them
3042 // scalar.
3043 SmallVector<VectorParts, 4> OpsV;
3044 for (Value *Op : Gep->operands()) {
3045 Instruction *SrcInst = dyn_cast<Instruction>(Op);
3046 if (SrcInst && OrigLoop->contains(SrcInst))
3047 OpsV.push_back(getVectorValue(Op));
3048 else
3049 OpsV.push_back(VectorParts(UF, Op));
3050 }
3051 for (unsigned Part = 0; Part < UF; ++Part) {
3052 SmallVector<Value *, 4> Ops;
3053 Value *GEPBasePtr = OpsV[0][Part];
3054 for (unsigned i = 1; i < Gep->getNumOperands(); i++)
3055 Ops.push_back(OpsV[i][Part]);
3056 Value *NewGep = Builder.CreateGEP(GEPBasePtr, Ops, "VectorGep");
3057 cast<GetElementPtrInst>(NewGep)->setIsInBounds(Gep->isInBounds());
3058 assert(NewGep->getType()->isVectorTy() && "Expected vector GEP")((NewGep->getType()->isVectorTy() && "Expected vector GEP"
) ? static_cast<void> (0) : __assert_fail ("NewGep->getType()->isVectorTy() && \"Expected vector GEP\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3058, __PRETTY_FUNCTION__))
;
3059
3060 NewGep =
3061 Builder.CreateBitCast(NewGep, VectorType::get(Ptr->getType(), VF));
3062 VectorGep.push_back(NewGep);
3063 }
3064 } else
3065 VectorGep = getVectorValue(Ptr);
3066 }
3067
3068 VectorParts Mask = createBlockInMask(Instr->getParent());
3069 // Handle Stores:
3070 if (SI) {
3071 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3072, __PRETTY_FUNCTION__))
3072 "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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3072, __PRETTY_FUNCTION__))
;
3073 setDebugLocFromInst(Builder, SI);
3074 // We don't want to update the value in the map as it might be used in
3075 // another expression. So don't use a reference type for "StoredVal".
3076 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
3077
3078 for (unsigned Part = 0; Part < UF; ++Part) {
3079 Instruction *NewSI = nullptr;
3080 if (CreateGatherScatter) {
3081 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
3082 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
3083 Alignment, MaskPart);
3084 } else {
3085 // Calculate the pointer for the specific unroll-part.
3086 Value *PartPtr =
3087 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3088
3089 if (Reverse) {
3090 // If we store to reverse consecutive memory locations, then we need
3091 // to reverse the order of elements in the stored value.
3092 StoredVal[Part] = reverseVector(StoredVal[Part]);
3093 // If the address is consecutive but reversed, then the
3094 // wide store needs to start at the last vector element.
3095 PartPtr =
3096 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3097 PartPtr =
3098 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3099 Mask[Part] = reverseVector(Mask[Part]);
3100 }
3101
3102 Value *VecPtr =
3103 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3104
3105 if (Legal->isMaskRequired(SI))
3106 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
3107 Mask[Part]);
3108 else
3109 NewSI =
3110 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
3111 }
3112 addMetadata(NewSI, SI);
3113 }
3114 return;
3115 }
3116
3117 // Handle loads.
3118 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3118, __PRETTY_FUNCTION__))
;
3119 setDebugLocFromInst(Builder, LI);
3120 VectorParts Entry(UF);
3121 for (unsigned Part = 0; Part < UF; ++Part) {
3122 Instruction *NewLI;
3123 if (CreateGatherScatter) {
3124 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
3125 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
3126 0, "wide.masked.gather");
3127 Entry[Part] = NewLI;
3128 } else {
3129 // Calculate the pointer for the specific unroll-part.
3130 Value *PartPtr =
3131 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3132
3133 if (Reverse) {
3134 // If the address is consecutive but reversed, then the
3135 // wide load needs to start at the last vector element.
3136 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3137 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3138 Mask[Part] = reverseVector(Mask[Part]);
3139 }
3140
3141 Value *VecPtr =
3142 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3143 if (Legal->isMaskRequired(LI))
3144 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
3145 UndefValue::get(DataTy),
3146 "wide.masked.load");
3147 else
3148 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
3149 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
3150 }
3151 addMetadata(NewLI, LI);
3152 }
3153 VectorLoopValueMap.initVector(Instr, Entry);
3154}
3155
3156void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
3157 bool IfPredicateInstr) {
3158 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3158, __PRETTY_FUNCTION__))
;
3159 DEBUG(dbgs() << "LV: Scalarizing"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
3160 << (IfPredicateInstr ? " and predicating:" : ":") << *Instrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
3161 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
;
3162 // Holds vector parameters or scalars, in case of uniform vals.
3163 SmallVector<VectorParts, 4> Params;
3164
3165 setDebugLocFromInst(Builder, Instr);
3166
3167 // Does this instruction return a value ?
3168 bool IsVoidRetTy = Instr->getType()->isVoidTy();
3169
3170 // Initialize a new scalar map entry.
3171 ScalarParts Entry(UF);
3172
3173 VectorParts Cond;
3174 if (IfPredicateInstr)
3175 Cond = createBlockInMask(Instr->getParent());
3176
3177 // Determine the number of scalars we need to generate for each unroll
3178 // iteration. If the instruction is uniform, we only need to generate the
3179 // first lane. Otherwise, we generate all VF values.
3180 unsigned Lanes = Cost->isUniformAfterVectorization(Instr, VF) ? 1 : VF;
3181
3182 // For each vector unroll 'part':
3183 for (unsigned Part = 0; Part < UF; ++Part) {
3184 Entry[Part].resize(VF);
3185 // For each scalar that we create:
3186 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
3187
3188 // Start if-block.
3189 Value *Cmp = nullptr;
3190 if (IfPredicateInstr) {
3191 Cmp = Cond[Part];
3192 if (Cmp->getType()->isVectorTy())
3193 Cmp = Builder.CreateExtractElement(Cmp, Builder.getInt32(Lane));
3194 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
3195 ConstantInt::get(Cmp->getType(), 1));
3196 }
3197
3198 Instruction *Cloned = Instr->clone();
3199 if (!IsVoidRetTy)
3200 Cloned->setName(Instr->getName() + ".cloned");
3201
3202 // Replace the operands of the cloned instructions with their scalar
3203 // equivalents in the new loop.
3204 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3205 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, Lane);
3206 Cloned->setOperand(op, NewOp);
3207 }
3208 addNewMetadata(Cloned, Instr);
3209
3210 // Place the cloned scalar in the new loop.
3211 Builder.Insert(Cloned);
3212
3213 // Add the cloned scalar to the scalar map entry.
3214 Entry[Part][Lane] = Cloned;
3215
3216 // If we just cloned a new assumption, add it the assumption cache.
3217 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3218 if (II->getIntrinsicID() == Intrinsic::assume)
3219 AC->registerAssumption(II);
3220
3221 // End if-block.
3222 if (IfPredicateInstr)
3223 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
3224 }
3225 }
3226 VectorLoopValueMap.initScalar(Instr, Entry);
3227}
3228
3229PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
3230 Value *End, Value *Step,
3231 Instruction *DL) {
3232 BasicBlock *Header = L->getHeader();
3233 BasicBlock *Latch = L->getLoopLatch();
3234 // As we're just creating this loop, it's possible no latch exists
3235 // yet. If so, use the header as this will be a single block loop.
3236 if (!Latch)
3237 Latch = Header;
3238
3239 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
3240 Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction);
3241 setDebugLocFromInst(Builder, OldInst);
3242 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3243
3244 Builder.SetInsertPoint(Latch->getTerminator());
3245 setDebugLocFromInst(Builder, OldInst);
3246
3247 // Create i+1 and fill the PHINode.
3248 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3249 Induction->addIncoming(Start, L->getLoopPreheader());
3250 Induction->addIncoming(Next, Latch);
3251 // Create the compare.
3252 Value *ICmp = Builder.CreateICmpEQ(Next, End);
3253 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3254
3255 // Now we have two terminators. Remove the old one from the block.
3256 Latch->getTerminator()->eraseFromParent();
3257
3258 return Induction;
3259}
3260
3261Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
3262 if (TripCount)
3263 return TripCount;
3264
3265 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3266 // Find the loop boundaries.
3267 ScalarEvolution *SE = PSE.getSE();
3268 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3269 assert(BackedgeTakenCount != SE->getCouldNotCompute() &&((BackedgeTakenCount != SE->getCouldNotCompute() &&
"Invalid loop count") ? static_cast<void> (0) : __assert_fail
("BackedgeTakenCount != SE->getCouldNotCompute() && \"Invalid loop count\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3270, __PRETTY_FUNCTION__))
3270 "Invalid loop count")((BackedgeTakenCount != SE->getCouldNotCompute() &&
"Invalid loop count") ? static_cast<void> (0) : __assert_fail
("BackedgeTakenCount != SE->getCouldNotCompute() && \"Invalid loop count\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3270, __PRETTY_FUNCTION__))
;
3271
3272 Type *IdxTy = Legal->getWidestInductionType();
3273
3274 // The exit count might have the type of i64 while the phi is i32. This can
3275 // happen if we have an induction variable that is sign extended before the
3276 // compare. The only way that we get a backedge taken count is that the
3277 // induction variable was signed and as such will not overflow. In such a case
3278 // truncation is legal.
3279 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3280 IdxTy->getPrimitiveSizeInBits())
3281 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3282 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3283
3284 // Get the total trip count from the count by adding 1.
3285 const SCEV *ExitCount = SE->getAddExpr(
3286 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3287
3288 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3289
3290 // Expand the trip count and place the new instructions in the preheader.
3291 // Notice that the pre-header does not change, only the loop body.
3292 SCEVExpander Exp(*SE, DL, "induction");
3293
3294 // Count holds the overall loop count (N).
3295 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3296 L->getLoopPreheader()->getTerminator());
3297
3298 if (TripCount->getType()->isPointerTy())
3299 TripCount =
3300 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3301 L->getLoopPreheader()->getTerminator());
3302
3303 return TripCount;
3304}
3305
3306Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
3307 if (VectorTripCount)
3308 return VectorTripCount;
3309
3310 Value *TC = getOrCreateTripCount(L);
3311 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3312
3313 // Now we need to generate the expression for the part of the loop that the
3314 // vectorized body will execute. This is equal to N - (N % Step) if scalar
3315 // iterations are not required for correctness, or N - Step, otherwise. Step
3316 // is equal to the vectorization factor (number of SIMD elements) times the
3317 // unroll factor (number of SIMD instructions).
3318 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3319 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3320
3321 // If there is a non-reversed interleaved group that may speculatively access
3322 // memory out-of-bounds, we need to ensure that there will be at least one
3323 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3324 // the trip count, we set the remainder to be equal to the step. If the step
3325 // does not evenly divide the trip count, no adjustment is necessary since
3326 // there will already be scalar iterations. Note that the minimum iterations
3327 // check ensures that N >= Step.
3328 if (VF > 1 && Legal->requiresScalarEpilogue()) {
3329 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3330 R = Builder.CreateSelect(IsZero, Step, R);
3331 }
3332
3333 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3334
3335 return VectorTripCount;
3336}
3337
3338void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
3339 BasicBlock *Bypass) {
3340 Value *Count = getOrCreateTripCount(L);
3341 BasicBlock *BB = L->getLoopPreheader();
3342 IRBuilder<> Builder(BB->getTerminator());
3343
3344 // Generate code to check that the loop's trip count that we computed by
3345 // adding one to the backedge-taken count will not overflow.
3346 Value *CheckMinIters = Builder.CreateICmpULT(
3347 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3348
3349 BasicBlock *NewBB =
3350 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
3351 // Update dominator tree immediately if the generated block is a
3352 // LoopBypassBlock because SCEV expansions to generate loop bypass
3353 // checks may query it before the current function is finished.
3354 DT->addNewBlock(NewBB, BB);
3355 if (L->getParentLoop())
3356 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3357 ReplaceInstWithInst(BB->getTerminator(),
3358 BranchInst::Create(Bypass, NewBB, CheckMinIters));
3359 LoopBypassBlocks.push_back(BB);
3360}
3361
3362void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
3363 BasicBlock *Bypass) {
3364 Value *TC = getOrCreateVectorTripCount(L);
3365 BasicBlock *BB = L->getLoopPreheader();
3366 IRBuilder<> Builder(BB->getTerminator());
3367
3368 // Now, compare the new count to zero. If it is zero skip the vector loop and
3369 // jump to the scalar loop.
3370 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
3371 "cmp.zero");
3372
3373 // Generate code to check that the loop's trip count that we computed by
3374 // adding one to the backedge-taken count will not overflow.
3375 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3376 // Update dominator tree immediately if the generated block is a
3377 // LoopBypassBlock because SCEV expansions to generate loop bypass
3378 // checks may query it before the current function is finished.
3379 DT->addNewBlock(NewBB, BB);
3380 if (L->getParentLoop())
3381 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3382 ReplaceInstWithInst(BB->getTerminator(),
3383 BranchInst::Create(Bypass, NewBB, Cmp));
3384 LoopBypassBlocks.push_back(BB);
3385}
3386
3387void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
3388 BasicBlock *BB = L->getLoopPreheader();
3389
3390 // Generate the code to check that the SCEV assumptions that we made.
3391 // We want the new basic block to start at the first instruction in a
3392 // sequence of instructions that form a check.
3393 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3394 "scev.check");
3395 Value *SCEVCheck =
3396 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3397
3398 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3399 if (C->isZero())
3400 return;
3401
3402 // Create a new block containing the stride check.
3403 BB->setName("vector.scevcheck");
3404 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3405 // Update dominator tree immediately if the generated block is a
3406 // LoopBypassBlock because SCEV expansions to generate loop bypass
3407 // checks may query it before the current function is finished.
3408 DT->addNewBlock(NewBB, BB);
3409 if (L->getParentLoop())
3410 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3411 ReplaceInstWithInst(BB->getTerminator(),
3412 BranchInst::Create(Bypass, NewBB, SCEVCheck));
3413 LoopBypassBlocks.push_back(BB);
3414 AddedSafetyChecks = true;
3415}
3416
3417void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
3418 BasicBlock *BB = L->getLoopPreheader();
3419
3420 // Generate the code that checks in runtime if arrays overlap. We put the
3421 // checks into a separate block to make the more common case of few elements
3422 // faster.
3423 Instruction *FirstCheckInst;
3424 Instruction *MemRuntimeCheck;
3425 std::tie(FirstCheckInst, MemRuntimeCheck) =
3426 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3427 if (!MemRuntimeCheck)
3428 return;
3429
3430 // Create a new block containing the memory check.
3431 BB->setName("vector.memcheck");
3432 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3433 // Update dominator tree immediately if the generated block is a
3434 // LoopBypassBlock because SCEV expansions to generate loop bypass
3435 // checks may query it before the current function is finished.
3436 DT->addNewBlock(NewBB, BB);
3437 if (L->getParentLoop())
3438 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3439 ReplaceInstWithInst(BB->getTerminator(),
3440 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3441 LoopBypassBlocks.push_back(BB);
3442 AddedSafetyChecks = true;
3443
3444 // We currently don't use LoopVersioning for the actual loop cloning but we
3445 // still use it to add the noalias metadata.
3446 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3447 PSE.getSE());
3448 LVer->prepareNoAliasMetadata();
3449}
3450
3451void InnerLoopVectorizer::createEmptyLoop() {
3452 /*
3453 In this function we generate a new loop. The new loop will contain
3454 the vectorized instructions while the old loop will continue to run the
3455 scalar remainder.
3456
3457 [ ] <-- loop iteration number check.
3458 / |
3459 / v
3460 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3461 | / |
3462 | / v
3463 || [ ] <-- vector pre header.
3464 |/ |
3465 | v
3466 | [ ] \
3467 | [ ]_| <-- vector loop.
3468 | |
3469 | v
3470 | -[ ] <--- middle-block.
3471 | / |
3472 | / v
3473 -|- >[ ] <--- new preheader.
3474 | |
3475 | v
3476 | [ ] \
3477 | [ ]_| <-- old scalar loop to handle remainder.
3478 \ |
3479 \ v
3480 >[ ] <-- exit block.
3481 ...
3482 */
3483
3484 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3485 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3486 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3487 assert(VectorPH && "Invalid loop structure")((VectorPH && "Invalid loop structure") ? static_cast
<void> (0) : __assert_fail ("VectorPH && \"Invalid loop structure\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3487, __PRETTY_FUNCTION__))
;
3488 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3488, __PRETTY_FUNCTION__))
;
3489
3490 // Some loops have a single integer induction variable, while other loops
3491 // don't. One example is c++ iterators that often have multiple pointer
3492 // induction variables. In the code below we also support a case where we
3493 // don't have a single induction variable.
3494 //
3495 // We try to obtain an induction variable from the original loop as hard
3496 // as possible. However if we don't find one that:
3497 // - is an integer
3498 // - counts from zero, stepping by one
3499 // - is the size of the widest induction variable type
3500 // then we create a new one.
3501 OldInduction = Legal->getPrimaryInduction();
3502 Type *IdxTy = Legal->getWidestInductionType();
3503
3504 // Split the single block loop into the two loop structure described above.
3505 BasicBlock *VecBody =
3506 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3507 BasicBlock *MiddleBlock =
3508 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3509 BasicBlock *ScalarPH =
3510 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3511
3512 // Create and register the new vector loop.
3513 Loop *Lp = new Loop();
3514 Loop *ParentLoop = OrigLoop->getParentLoop();
3515
3516 // Insert the new loop into the loop nest and register the new basic blocks
3517 // before calling any utilities such as SCEV that require valid LoopInfo.
3518 if (ParentLoop) {
3519 ParentLoop->addChildLoop(Lp);
3520 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3521 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3522 } else {
3523 LI->addTopLevelLoop(Lp);
3524 }
3525 Lp->addBasicBlockToLoop(VecBody, *LI);
3526
3527 // Find the loop boundaries.
3528 Value *Count = getOrCreateTripCount(Lp);
3529
3530 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3531
3532 // We need to test whether the backedge-taken count is uint##_max. Adding one
3533 // to it will cause overflow and an incorrect loop trip count in the vector
3534 // body. In case of overflow we want to directly jump to the scalar remainder
3535 // loop.
3536 emitMinimumIterationCountCheck(Lp, ScalarPH);
3537 // Now, compare the new count to zero. If it is zero skip the vector loop and
3538 // jump to the scalar loop.
3539 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3540 // Generate the code to check any assumptions that we've made for SCEV
3541 // expressions.
3542 emitSCEVChecks(Lp, ScalarPH);
3543
3544 // Generate the code that checks in runtime if arrays overlap. We put the
3545 // checks into a separate block to make the more common case of few elements
3546 // faster.
3547 emitMemRuntimeChecks(Lp, ScalarPH);
3548
3549 // Generate the induction variable.
3550 // The loop step is equal to the vectorization factor (num of SIMD elements)
3551 // times the unroll factor (num of SIMD instructions).
3552 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3553 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3554 Induction =
3555 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3556 getDebugLocFromInstOrOperands(OldInduction));
3557
3558 // We are going to resume the execution of the scalar loop.
3559 // Go over all of the induction variables that we found and fix the
3560 // PHIs that are left in the scalar version of the loop.
3561 // The starting values of PHI nodes depend on the counter of the last
3562 // iteration in the vectorized loop.
3563 // If we come from a bypass edge then we need to start from the original
3564 // start value.
3565
3566 // This variable saves the new starting index for the scalar loop. It is used
3567 // to test if there are any tail iterations left once the vector loop has
3568 // completed.
3569 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3570 for (auto &InductionEntry : *List) {
3571 PHINode *OrigPhi = InductionEntry.first;
3572 InductionDescriptor II = InductionEntry.second;
3573
3574 // Create phi nodes to merge from the backedge-taken check block.
3575 PHINode *BCResumeVal = PHINode::Create(
3576 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3577 Value *&EndValue = IVEndValues[OrigPhi];
3578 if (OrigPhi == OldInduction) {
3579 // We know what the end value is.
3580 EndValue = CountRoundDown;
3581 } else {
3582 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3583 Type *StepType = II.getStep()->getType();
3584 Instruction::CastOps CastOp =
3585 CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3586 Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3587 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3588 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3589 EndValue->setName("ind.end");
3590 }
3591
3592 // The new PHI merges the original incoming value, in case of a bypass,
3593 // or the value at the end of the vectorized loop.
3594 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3595
3596 // Fix the scalar body counter (PHI node).
3597 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3598
3599 // The old induction's phi node in the scalar body needs the truncated
3600 // value.
3601 for (BasicBlock *BB : LoopBypassBlocks)
3602 BCResumeVal->addIncoming(II.getStartValue(), BB);
3603 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3604 }
3605
3606 // Add a check in the middle block to see if we have completed
3607 // all of the iterations in the first vector loop.
3608 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3609 Value *CmpN =
3610 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3611 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3612 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3613 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3614
3615 // Get ready to start creating new instructions into the vectorized body.
3616 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3617
3618 // Save the state.
3619 LoopVectorPreHeader = Lp->getLoopPreheader();
3620 LoopScalarPreHeader = ScalarPH;
3621 LoopMiddleBlock = MiddleBlock;
3622 LoopExitBlock = ExitBlock;
3623 LoopVectorBody = VecBody;
3624 LoopScalarBody = OldBasicBlock;
3625
3626 // Keep all loop hints from the original loop on the vector loop (we'll
3627 // replace the vectorizer-specific hints below).
3628 if (MDNode *LID = OrigLoop->getLoopID())
3629 Lp->setLoopID(LID);
3630
3631 LoopVectorizeHints Hints(Lp, true, *ORE);
3632 Hints.setAlreadyVectorized();
3633}
3634
3635// Fix up external users of the induction variable. At this point, we are
3636// in LCSSA form, with all external PHIs that use the IV having one input value,
3637// coming from the remainder loop. We need those PHIs to also have a correct
3638// value for the IV when arriving directly from the middle block.
3639void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3640 const InductionDescriptor &II,
3641 Value *CountRoundDown, Value *EndValue,
3642 BasicBlock *MiddleBlock) {
3643 // There are two kinds of external IV usages - those that use the value
3644 // computed in the last iteration (the PHI) and those that use the penultimate
3645 // value (the value that feeds into the phi from the loop latch).
3646 // We allow both, but they, obviously, have different values.
3647
3648 assert(OrigLoop->getExitBlock() && "Expected a single exit block")((OrigLoop->getExitBlock() && "Expected a single exit block"
) ? static_cast<void> (0) : __assert_fail ("OrigLoop->getExitBlock() && \"Expected a single exit block\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3648, __PRETTY_FUNCTION__))
;
3649
3650 DenseMap<Value *, Value *> MissingVals;
3651
3652 // An external user of the last iteration's value should see the value that
3653 // the remainder loop uses to initialize its own IV.
3654 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3655 for (User *U : PostInc->users()) {
3656 Instruction *UI = cast<Instruction>(U);
3657 if (!OrigLoop->contains(UI)) {
3658 assert(isa<PHINode>(UI) && "Expected LCSSA form")((isa<PHINode>(UI) && "Expected LCSSA form") ? static_cast
<void> (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3658, __PRETTY_FUNCTION__))
;
3659 MissingVals[UI] = EndValue;
3660 }
3661 }
3662
3663 // An external user of the penultimate value need to see EndValue - Step.
3664 // The simplest way to get this is to recompute it from the constituent SCEVs,
3665 // that is Start + (Step * (CRD - 1)).
3666 for (User *U : OrigPhi->users()) {
3667 auto *UI = cast<Instruction>(U);
3668 if (!OrigLoop->contains(UI)) {
3669 const DataLayout &DL =
3670 OrigLoop->getHeader()->getModule()->getDataLayout();
3671 assert(isa<PHINode>(UI) && "Expected LCSSA form")((isa<PHINode>(UI) && "Expected LCSSA form") ? static_cast
<void> (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3671, __PRETTY_FUNCTION__))
;
3672
3673 IRBuilder<> B(MiddleBlock->getTerminator());
3674 Value *CountMinusOne = B.CreateSub(
3675 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3676 Value *CMO = B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType(),
3677 "cast.cmo");
3678 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3679 Escape->setName("ind.escape");
3680 MissingVals[UI] = Escape;
3681 }
3682 }
3683
3684 for (auto &I : MissingVals) {
3685 PHINode *PHI = cast<PHINode>(I.first);
3686 // One corner case we have to handle is two IVs "chasing" each-other,
3687 // that is %IV2 = phi [...], [ %IV1, %latch ]
3688 // In this case, if IV1 has an external use, we need to avoid adding both
3689 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3690 // don't already have an incoming value for the middle block.
3691 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3692 PHI->addIncoming(I.second, MiddleBlock);
3693 }
3694}
3695
3696namespace {
3697struct CSEDenseMapInfo {
3698 static bool canHandle(const Instruction *I) {
3699 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3700 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3701 }
3702 static inline Instruction *getEmptyKey() {
3703 return DenseMapInfo<Instruction *>::getEmptyKey();
3704 }
3705 static inline Instruction *getTombstoneKey() {
3706 return DenseMapInfo<Instruction *>::getTombstoneKey();
3707 }
3708 static unsigned getHashValue(const Instruction *I) {
3709 assert(canHandle(I) && "Unknown instruction!")((canHandle(I) && "Unknown instruction!") ? static_cast
<void> (0) : __assert_fail ("canHandle(I) && \"Unknown instruction!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3709, __PRETTY_FUNCTION__))
;
3710 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3711 I->value_op_end()));
3712 }
3713 static bool isEqual(const Instruction *LHS, const Instruction *RHS) {
3714 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3715 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3716 return LHS == RHS;
3717 return LHS->isIdenticalTo(RHS);
3718 }
3719};
3720}
3721
3722///\brief Perform cse of induction variable instructions.
3723static void cse(BasicBlock *BB) {
3724 // Perform simple cse.
3725 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3726 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3727 Instruction *In = &*I++;
3728
3729 if (!CSEDenseMapInfo::canHandle(In))
3730 continue;
3731
3732 // Check if we can replace this instruction with any of the
3733 // visited instructions.
3734 if (Instruction *V = CSEMap.lookup(In)) {
3735 In->replaceAllUsesWith(V);
3736 In->eraseFromParent();
3737 continue;
3738 }
3739
3740 CSEMap[In] = In;
3741 }
3742}
3743
3744/// \brief Estimate the overhead of scalarizing an instruction. This is a
3745/// convenience wrapper for the type-based getScalarizationOverhead API.
3746static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3747 const TargetTransformInfo &TTI) {
3748 if (VF == 1)
3749 return 0;
3750
3751 unsigned Cost = 0;
3752 Type *RetTy = ToVectorTy(I->getType(), VF);
3753 if (!RetTy->isVoidTy())
3754 Cost += TTI.getScalarizationOverhead(RetTy, true, false);
3755
3756 if (CallInst *CI = dyn_cast<CallInst>(I)) {
3757 SmallVector<const Value *, 4> Operands(CI->arg_operands());
3758 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3759 } else {
3760 SmallVector<const Value *, 4> Operands(I->operand_values());
3761 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3762 }
3763
3764 return Cost;
3765}
3766
3767// Estimate cost of a call instruction CI if it were vectorized with factor VF.
3768// Return the cost of the instruction, including scalarization overhead if it's
3769// needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3770// i.e. either vector version isn't available, or is too expensive.
3771static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3772 const TargetTransformInfo &TTI,
3773 const TargetLibraryInfo *TLI,
3774 bool &NeedToScalarize) {
3775 Function *F = CI->getCalledFunction();
3776 StringRef FnName = CI->getCalledFunction()->getName();
3777 Type *ScalarRetTy = CI->getType();
3778 SmallVector<Type *, 4> Tys, ScalarTys;
3779 for (auto &ArgOp : CI->arg_operands())
3780 ScalarTys.push_back(ArgOp->getType());
3781
3782 // Estimate cost of scalarized vector call. The source operands are assumed
3783 // to be vectors, so we need to extract individual elements from there,
3784 // execute VF scalar calls, and then gather the result into the vector return
3785 // value.
3786 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3787 if (VF == 1)
3788 return ScalarCallCost;
3789
3790 // Compute corresponding vector type for return value and arguments.
3791 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3792 for (Type *ScalarTy : ScalarTys)
3793 Tys.push_back(ToVectorTy(ScalarTy, VF));
3794
3795 // Compute costs of unpacking argument values for the scalar calls and
3796 // packing the return values to a vector.
3797 unsigned ScalarizationCost = getScalarizationOverhead(CI, VF, TTI);
3798
3799 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3800
3801 // If we can't emit a vector call for this function, then the currently found
3802 // cost is the cost we need to return.
3803 NeedToScalarize = true;
3804 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3805 return Cost;
3806
3807 // If the corresponding vector cost is cheaper, return its cost.
3808 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3809 if (VectorCallCost < Cost) {
3810 NeedToScalarize = false;
3811 return VectorCallCost;
3812 }
3813 return Cost;
3814}
3815
3816// Estimate cost of an intrinsic call instruction CI if it were vectorized with
3817// factor VF. Return the cost of the instruction, including scalarization
3818// overhead if it's needed.
3819static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3820 const TargetTransformInfo &TTI,
3821 const TargetLibraryInfo *TLI) {
3822 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3823 assert(ID && "Expected intrinsic call!")((ID && "Expected intrinsic call!") ? static_cast<
void> (0) : __assert_fail ("ID && \"Expected intrinsic call!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3823, __PRETTY_FUNCTION__))
;
3824
3825 FastMathFlags FMF;
3826 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3827 FMF = FPMO->getFastMathFlags();
3828
3829 SmallVector<Value *, 4> Operands(CI->arg_operands());
3830 return TTI.getIntrinsicInstrCost(ID, CI->getType(), Operands, FMF, VF);
3831}
3832
3833static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3834 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3835 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3836 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3837}
3838static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3839 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3840 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3841 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3842}
3843
3844void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3845 // For every instruction `I` in MinBWs, truncate the operands, create a
3846 // truncated version of `I` and reextend its result. InstCombine runs
3847 // later and will remove any ext/trunc pairs.
3848 //
3849 SmallPtrSet<Value *, 4> Erased;
3850 for (const auto &KV : Cost->getMinimalBitwidths()) {
3851 // If the value wasn't vectorized, we must maintain the original scalar
3852 // type. The absence of the value from VectorLoopValueMap indicates that it
3853 // wasn't vectorized.
3854 if (!VectorLoopValueMap.hasVector(KV.first))
3855 continue;
3856 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3857 for (Value *&I : Parts) {
3858 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3859 continue;
3860 Type *OriginalTy = I->getType();
3861 Type *ScalarTruncatedTy =
3862 IntegerType::get(OriginalTy->getContext(), KV.second);
3863 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3864 OriginalTy->getVectorNumElements());
3865 if (TruncatedTy == OriginalTy)
3866 continue;
3867
3868 IRBuilder<> B(cast<Instruction>(I));
3869 auto ShrinkOperand = [&](Value *V) -> Value * {
3870 if (auto *ZI = dyn_cast<ZExtInst>(V))
3871 if (ZI->getSrcTy() == TruncatedTy)
3872 return ZI->getOperand(0);
3873 return B.CreateZExtOrTrunc(V, TruncatedTy);
3874 };
3875
3876 // The actual instruction modification depends on the instruction type,
3877 // unfortunately.
3878 Value *NewI = nullptr;
3879 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3880 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3881 ShrinkOperand(BO->getOperand(1)));
3882 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3883 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3884 NewI =
3885 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3886 ShrinkOperand(CI->getOperand(1)));
3887 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3888 NewI = B.CreateSelect(SI->getCondition(),
3889 ShrinkOperand(SI->getTrueValue()),
3890 ShrinkOperand(SI->getFalseValue()));
3891 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3892 switch (CI->getOpcode()) {
3893 default:
3894 llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3894)
;
3895 case Instruction::Trunc:
3896 NewI = ShrinkOperand(CI->getOperand(0));
3897 break;
3898 case Instruction::SExt:
3899 NewI = B.CreateSExtOrTrunc(
3900 CI->getOperand(0),
3901 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3902 break;
3903 case Instruction::ZExt:
3904 NewI = B.CreateZExtOrTrunc(
3905 CI->getOperand(0),
3906 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3907 break;
3908 }
3909 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3910 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3911 auto *O0 = B.CreateZExtOrTrunc(
3912 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3913 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3914 auto *O1 = B.CreateZExtOrTrunc(
3915 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3916
3917 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3918 } else if (isa<LoadInst>(I)) {
3919 // Don't do anything with the operands, just extend the result.
3920 continue;
3921 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3922 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3923 auto *O0 = B.CreateZExtOrTrunc(
3924 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3925 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3926 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3927 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3928 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3929 auto *O0 = B.CreateZExtOrTrunc(
3930 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3931 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3932 } else {
3933 llvm_unreachable("Unhandled instruction type!")::llvm::llvm_unreachable_internal("Unhandled instruction type!"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3933)
;
3934 }
3935
3936 // Lastly, extend the result.
3937 NewI->takeName(cast<Instruction>(I));
3938 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3939 I->replaceAllUsesWith(Res);
3940 cast<Instruction>(I)->eraseFromParent();
3941 Erased.insert(I);
3942 I = Res;
3943 }
3944 }
3945
3946 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3947 for (const auto &KV : Cost->getMinimalBitwidths()) {
3948 // If the value wasn't vectorized, we must maintain the original scalar
3949 // type. The absence of the value from VectorLoopValueMap indicates that it
3950 // wasn't vectorized.
3951 if (!VectorLoopValueMap.hasVector(KV.first))
3952 continue;
3953 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3954 for (Value *&I : Parts) {
3955 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3956 if (Inst && Inst->use_empty()) {
3957 Value *NewI = Inst->getOperand(0);
3958 Inst->eraseFromParent();
3959 I = NewI;
3960 }
3961 }
3962 }
3963}
3964
3965void InnerLoopVectorizer::vectorizeLoop() {
3966 //===------------------------------------------------===//
3967 //
3968 // Notice: any optimization or new instruction that go
3969 // into the code below should be also be implemented in
3970 // the cost-model.
3971 //
3972 //===------------------------------------------------===//
3973
3974 // Collect instructions from the original loop that will become trivially
3975 // dead in the vectorized loop. We don't need to vectorize these
3976 // instructions.
3977 collectTriviallyDeadInstructions();
3978
3979 // Scan the loop in a topological order to ensure that defs are vectorized
3980 // before users.
3981 LoopBlocksDFS DFS(OrigLoop);
3982 DFS.perform(LI);
3983
3984 // Vectorize all of the blocks in the original loop.
3985 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
3986 vectorizeBlockInLoop(BB);
3987
3988 // Insert truncates and extends for any truncated instructions as hints to
3989 // InstCombine.
3990 if (VF > 1)
3991 truncateToMinimalBitwidths();
3992
3993 // At this point every instruction in the original loop is widened to a
3994 // vector form. Now we need to fix the recurrences in the loop. These PHI
3995 // nodes are currently empty because we did not want to introduce cycles.
3996 // This is the second stage of vectorizing recurrences.
3997 fixCrossIterationPHIs();
3998
3999 // Update the dominator tree.
4000 //
4001 // FIXME: After creating the structure of the new loop, the dominator tree is
4002 // no longer up-to-date, and it remains that way until we update it
4003 // here. An out-of-date dominator tree is problematic for SCEV,
4004 // because SCEVExpander uses it to guide code generation. The
4005 // vectorizer use SCEVExpanders in several places. Instead, we should
4006 // keep the dominator tree up-to-date as we go.
4007 updateAnalysis();
4008
4009 // Fix-up external users of the induction variables.
4010 for (auto &Entry : *Legal->getInductionVars())
4011 fixupIVUsers(Entry.first, Entry.second,
4012 getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)),
4013 IVEndValues[Entry.first], LoopMiddleBlock);
4014
4015 fixLCSSAPHIs();
4016 predicateInstructions();
4017
4018 // Remove redundant induction instructions.
4019 cse(LoopVectorBody);
4020}
4021
4022void InnerLoopVectorizer::fixCrossIterationPHIs() {
4023 // In order to support recurrences we need to be able to vectorize Phi nodes.
4024 // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4025 // stage #2: We now need to fix the recurrences by adding incoming edges to
4026 // the currently empty PHI nodes. At this point every instruction in the
4027 // original loop is widened to a vector form so we can use them to construct
4028 // the incoming edges.
4029 for (Instruction &I : *OrigLoop->getHeader()) {
4030 PHINode *Phi = dyn_cast<PHINode>(&I);
4031 if (!Phi)
4032 break;
4033 // Handle first-order recurrences and reductions that need to be fixed.
4034 if (Legal->isFirstOrderRecurrence(Phi))
4035 fixFirstOrderRecurrence(Phi);
4036 else if (Legal->isReductionVariable(Phi))
4037 fixReduction(Phi);
4038 }
4039}
4040
4041void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
4042
4043 // This is the second phase of vectorizing first-order recurrences. An
4044 // overview of the transformation is described below. Suppose we have the
4045 // following loop.
4046 //
4047 // for (int i = 0; i < n; ++i)
4048 // b[i] = a[i] - a[i - 1];
4049 //
4050 // There is a first-order recurrence on "a". For this loop, the shorthand
4051 // scalar IR looks like:
4052 //
4053 // scalar.ph:
4054 // s_init = a[-1]
4055 // br scalar.body
4056 //
4057 // scalar.body:
4058 // i = phi [0, scalar.ph], [i+1, scalar.body]
4059 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
4060 // s2 = a[i]
4061 // b[i] = s2 - s1
4062 // br cond, scalar.body, ...
4063 //
4064 // In this example, s1 is a recurrence because it's value depends on the
4065 // previous iteration. In the first phase of vectorization, we created a
4066 // temporary value for s1. We now complete the vectorization and produce the
4067 // shorthand vector IR shown below (for VF = 4, UF = 1).
4068 //
4069 // vector.ph:
4070 // v_init = vector(..., ..., ..., a[-1])
4071 // br vector.body
4072 //
4073 // vector.body
4074 // i = phi [0, vector.ph], [i+4, vector.body]
4075 // v1 = phi [v_init, vector.ph], [v2, vector.body]
4076 // v2 = a[i, i+1, i+2, i+3];
4077 // v3 = vector(v1(3), v2(0, 1, 2))
4078 // b[i, i+1, i+2, i+3] = v2 - v3
4079 // br cond, vector.body, middle.block
4080 //
4081 // middle.block:
4082 // x = v2(3)
4083 // br scalar.ph
4084 //
4085 // scalar.ph:
4086 // s_init = phi [x, middle.block], [a[-1], otherwise]
4087 // br scalar.body
4088 //
4089 // After execution completes the vector loop, we extract the next value of
4090 // the recurrence (x) to use as the initial value in the scalar loop.
4091
4092 // Get the original loop preheader and single loop latch.
4093 auto *Preheader = OrigLoop->getLoopPreheader();
4094 auto *Latch = OrigLoop->getLoopLatch();
4095
4096 // Get the initial and previous values of the scalar recurrence.
4097 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4098 auto *Previous = Phi->getIncomingValueForBlock(Latch);
4099
4100 // Create a vector from the initial value.
4101 auto *VectorInit = ScalarInit;
4102 if (VF > 1) {
4103 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
4104 VectorInit = Builder.CreateInsertElement(
4105 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4106 Builder.getInt32(VF - 1), "vector.recur.init");
4107 }
4108
4109 // We constructed a temporary phi node in the first phase of vectorization.
4110 // This phi node will eventually be deleted.
4111 VectorParts &PhiParts = VectorLoopValueMap.getVector(Phi);
4112 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
4113
4114 // Create a phi node for the new recurrence. The current value will either be
4115 // the initial value inserted into a vector or loop-varying vector value.
4116 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4117 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4118
4119 // Get the vectorized previous value.
4120 auto &PreviousParts = getVectorValue(Previous);
4121
4122 // Set the insertion point after the previous value if it is an instruction.
4123 // Note that the previous value may have been constant-folded so it is not
4124 // guaranteed to be an instruction in the vector loop.
4125 if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousParts[UF - 1]))
4126 Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt());
4127 else
4128 Builder.SetInsertPoint(
4129 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
4130
4131 // We will construct a vector for the recurrence by combining the values for
4132 // the current and previous iterations. This is the required shuffle mask.
4133 SmallVector<Constant *, 8> ShuffleMask(VF);
4134 ShuffleMask[0] = Builder.getInt32(VF - 1);
4135 for (unsigned I = 1; I < VF; ++I)
4136 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4137
4138 // The vector from which to take the initial value for the current iteration
4139 // (actual or unrolled). Initially, this is the vector phi node.
4140 Value *Incoming = VecPhi;
4141
4142 // Shuffle the current and previous vector and update the vector parts.
4143 for (unsigned Part = 0; Part < UF; ++Part) {
4144 auto *Shuffle =
4145 VF > 1
4146 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
4147 ConstantVector::get(ShuffleMask))
4148 : Incoming;
4149 PhiParts[Part]->replaceAllUsesWith(Shuffle);
4150 cast<Instruction>(PhiParts[Part])->eraseFromParent();
4151 PhiParts[Part] = Shuffle;
4152 Incoming = PreviousParts[Part];
4153 }
4154
4155 // Fix the latch value of the new recurrence in the vector loop.
4156 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4157
4158 // Extract the last vector element in the middle block. This will be the
4159 // initial value for the recurrence when jumping to the scalar loop.
4160 auto *Extract = Incoming;
4161 if (VF > 1) {
4162 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
4163 Extract = Builder.CreateExtractElement(Extract, Builder.getInt32(VF - 1),
4164 "vector.recur.extract");
4165 }
4166
4167 // Fix the initial value of the original recurrence in the scalar loop.
4168 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
4169 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4170 for (auto *BB : predecessors(LoopScalarPreHeader)) {
4171 auto *Incoming = BB == LoopMiddleBlock ? Extract : ScalarInit;
4172 Start->addIncoming(Incoming, BB);
4173 }
4174
4175 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
4176 Phi->setName("scalar.recur");
4177
4178 // Finally, fix users of the recurrence outside the loop. The users will need
4179 // either the last value of the scalar recurrence or the last value of the
4180 // vector recurrence we extracted in the middle block. Since the loop is in
4181 // LCSSA form, we just need to find the phi node for the original scalar
4182 // recurrence in the exit block, and then add an edge for the middle block.
4183 for (auto &I : *LoopExitBlock) {
4184 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4185 if (!LCSSAPhi)
4186 break;
4187 if (LCSSAPhi->getIncomingValue(0) == Phi) {
4188 LCSSAPhi->addIncoming(Extract, LoopMiddleBlock);
4189 break;
4190 }
4191 }
4192}
4193
4194void InnerLoopVectorizer::fixReduction(PHINode *Phi) {
4195 Constant *Zero = Builder.getInt32(0);
4196
4197 // Get it's reduction variable descriptor.
4198 assert(Legal->isReductionVariable(Phi) &&((Legal->isReductionVariable(Phi) && "Unable to find the reduction variable"
) ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(Phi) && \"Unable to find the reduction variable\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4199, __PRETTY_FUNCTION__))
4199 "Unable to find the reduction variable")((Legal->isReductionVariable(Phi) && "Unable to find the reduction variable"
) ? static_cast<void> (0) : __assert_fail ("Legal->isReductionVariable(Phi) && \"Unable to find the reduction variable\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4199, __PRETTY_FUNCTION__))
;
4200 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
4201
4202 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
4203 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
4204 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
4205 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
4206 RdxDesc.getMinMaxRecurrenceKind();
4207 setDebugLocFromInst(Builder, ReductionStartValue);
4208
4209 // We need to generate a reduction vector from the incoming scalar.
4210 // To do so, we need to generate the 'identity' vector and override
4211 // one of the elements with the incoming scalar reduction. We need
4212 // to do it in the vector-loop preheader.
4213 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
4214
4215 // This is the vector-clone of the value that leaves the loop.
4216 const VectorParts &VectorExit = getVectorValue(LoopExitInst);
4217 Type *VecTy = VectorExit[0]->getType();
4218
4219 // Find the reduction identity variable. Zero for addition, or, xor,
4220 // one for multiplication, -1 for And.
4221 Value *Identity;
4222 Value *VectorStart;
4223 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
4224 RK == RecurrenceDescriptor::RK_FloatMinMax) {
4225 // MinMax reduction have the start value as their identify.
4226 if (VF == 1) {
4227 VectorStart = Identity = ReductionStartValue;
4228 } else {
4229 VectorStart = Identity =
4230 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
4231 }
4232 } else {
4233 // Handle other reduction kinds:
4234 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
4235 RK, VecTy->getScalarType());
4236 if (VF == 1) {
4237 Identity = Iden;
4238 // This vector is the Identity vector where the first element is the
4239 // incoming scalar reduction.
4240 VectorStart = ReductionStartValue;
4241 } else {
4242 Identity = ConstantVector::getSplat(VF, Iden);
4243
4244 // This vector is the Identity vector where the first element is the
4245 // incoming scalar reduction.
4246 VectorStart =
4247 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
4248 }
4249 }
4250
4251 // Fix the vector-loop phi.
4252
4253 // Reductions do not have to start at zero. They can start with
4254 // any loop invariant values.
4255 const VectorParts &VecRdxPhi = getVectorValue(Phi);
4256 BasicBlock *Latch = OrigLoop->getLoopLatch();
4257 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
4258 const VectorParts &Val = getVectorValue(LoopVal);
4259 for (unsigned part = 0; part < UF; ++part) {
4260 // Make sure to add the reduction stat value only to the
4261 // first unroll part.
4262 Value *StartVal = (part == 0) ? VectorStart : Identity;
4263 cast<PHINode>(VecRdxPhi[part])
4264 ->addIncoming(StartVal, LoopVectorPreHeader);
4265 cast<PHINode>(VecRdxPhi[part])
4266 ->addIncoming(Val[part], LoopVectorBody);
4267 }
4268
4269 // Before each round, move the insertion point right between
4270 // the PHIs and the values we are going to write.
4271 // This allows us to write both PHINodes and the extractelement
4272 // instructions.
4273 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4274
4275 VectorParts &RdxParts = VectorLoopValueMap.getVector(LoopExitInst);
4276 setDebugLocFromInst(Builder, LoopExitInst);
4277
4278 // If the vector reduction can be performed in a smaller type, we truncate
4279 // then extend the loop exit value to enable InstCombine to evaluate the
4280 // entire expression in the smaller type.
4281 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
4282 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
4283 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
4284 for (unsigned part = 0; part < UF; ++part) {
4285 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4286 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4287 : Builder.CreateZExt(Trunc, VecTy);
4288 for (Value::user_iterator UI = RdxParts[part]->user_begin();
4289 UI != RdxParts[part]->user_end();)
4290 if (*UI != Trunc) {
4291 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
4292 RdxParts[part] = Extnd;
4293 } else {
4294 ++UI;
4295 }
4296 }
4297 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4298 for (unsigned part = 0; part < UF; ++part)
4299 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4300 }
4301
4302 // Reduce all of the unrolled parts into a single vector.
4303 Value *ReducedPartRdx = RdxParts[0];
4304 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
4305 setDebugLocFromInst(Builder, ReducedPartRdx);
4306 for (unsigned part = 1; part < UF; ++part) {
4307 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4308 // Floating point operations had to be 'fast' to enable the reduction.
4309 ReducedPartRdx = addFastMathFlag(
4310 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
4311 ReducedPartRdx, "bin.rdx"));
4312 else
4313 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4314 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
4315 }
4316
4317 if (VF > 1) {
4318 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
4319 // and vector ops, reducing the set of values being computed by half each
4320 // round.
4321 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4322, __PRETTY_FUNCTION__))
4322 "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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4322, __PRETTY_FUNCTION__))
;
4323 Value *TmpVec = ReducedPartRdx;
4324 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
4325 for (unsigned i = VF; i != 1; i >>= 1) {
4326 // Move the upper half of the vector to the lower half.
4327 for (unsigned j = 0; j != i / 2; ++j)
4328 ShuffleMask[j] = Builder.getInt32(i / 2 + j);
4329
4330 // Fill the rest of the mask with undef.
4331 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
4332 UndefValue::get(Builder.getInt32Ty()));
4333
4334 Value *Shuf = Builder.CreateShuffleVector(
4335 TmpVec, UndefValue::get(TmpVec->getType()),
4336 ConstantVector::get(ShuffleMask), "rdx.shuf");
4337
4338 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4339 // Floating point operations had to be 'fast' to enable the reduction.
4340 TmpVec = addFastMathFlag(Builder.CreateBinOp(
4341 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
4342 else
4343 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
4344 TmpVec, Shuf);
4345 }
4346
4347 // The result is in the first element of the vector.
4348 ReducedPartRdx =
4349 Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4350
4351 // If the reduction can be performed in a smaller type, we need to extend
4352 // the reduction to the wider type before we branch to the original loop.
4353 if (Phi->getType() != RdxDesc.getRecurrenceType())
4354 ReducedPartRdx =
4355 RdxDesc.isSigned()
4356 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4357 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4358 }
4359
4360 // Create a phi node that merges control-flow from the backedge-taken check
4361 // block and the middle block.
4362 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4363 LoopScalarPreHeader->getTerminator());
4364 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4365 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4366 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4367
4368 // Now, we need to fix the users of the reduction variable
4369 // inside and outside of the scalar remainder loop.
4370 // We know that the loop is in LCSSA form. We need to update the
4371 // PHI nodes in the exit blocks.
4372 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4373 LEE = LoopExitBlock->end();
4374 LEI != LEE; ++LEI) {
4375 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4376 if (!LCSSAPhi)
4377 break;
4378
4379 // All PHINodes need to have a single entry edge, or two if
4380 // we already fixed them.
4381 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4381, __PRETTY_FUNCTION__))
;
4382
4383 // We found a reduction value exit-PHI. Update it with the
4384 // incoming bypass edge.
4385 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst)
4386 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4387 } // end of the LCSSA phi scan.
4388
4389 // Fix the scalar loop reduction variable with the incoming reduction sum
4390 // from the vector body and from the backedge value.
4391 int IncomingEdgeBlockIdx =
4392 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
4393 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4393, __PRETTY_FUNCTION__))
;
4394 // Pick the other block.
4395 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4396 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4397 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4398}
4399
4400void InnerLoopVectorizer::fixLCSSAPHIs() {
4401 for (Instruction &LEI : *LoopExitBlock) {
4402 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4403 if (!LCSSAPhi)
4404 break;
4405 if (LCSSAPhi->getNumIncomingValues() == 1)
4406 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
4407 LoopMiddleBlock);
4408 }
4409}
4410
4411void InnerLoopVectorizer::collectTriviallyDeadInstructions() {
4412 BasicBlock *Latch = OrigLoop->getLoopLatch();
4413
4414 // We create new control-flow for the vectorized loop, so the original
4415 // condition will be dead after vectorization if it's only used by the
4416 // branch.
4417 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
4418 if (Cmp && Cmp->hasOneUse())
4419 DeadInstructions.insert(Cmp);
4420
4421 // We create new "steps" for induction variable updates to which the original
4422 // induction variables map. An original update instruction will be dead if
4423 // all its users except the induction variable are dead.
4424 for (auto &Induction : *Legal->getInductionVars()) {
4425 PHINode *Ind = Induction.first;
4426 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
4427 if (all_of(IndUpdate->users(), [&](User *U) -> bool {
4428 return U == Ind || DeadInstructions.count(cast<Instruction>(U));
4429 }))
4430 DeadInstructions.insert(IndUpdate);
4431 }
4432}
4433
4434void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) {
4435
4436 // The basic block and loop containing the predicated instruction.
4437 auto *PredBB = PredInst->getParent();
4438 auto *VectorLoop = LI->getLoopFor(PredBB);
4439
4440 // Initialize a worklist with the operands of the predicated instruction.
4441 SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4442
4443 // Holds instructions that we need to analyze again. An instruction may be
4444 // reanalyzed if we don't yet know if we can sink it or not.
4445 SmallVector<Instruction *, 8> InstsToReanalyze;
4446
4447 // Returns true if a given use occurs in the predicated block. Phi nodes use
4448 // their operands in their corresponding predecessor blocks.
4449 auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4450 auto *I = cast<Instruction>(U.getUser());
4451 BasicBlock *BB = I->getParent();
4452 if (auto *Phi = dyn_cast<PHINode>(I))
4453 BB = Phi->getIncomingBlock(
4454 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4455 return BB == PredBB;
4456 };
4457
4458 // Iteratively sink the scalarized operands of the predicated instruction
4459 // into the block we created for it. When an instruction is sunk, it's
4460 // operands are then added to the worklist. The algorithm ends after one pass
4461 // through the worklist doesn't sink a single instruction.
4462 bool Changed;
4463 do {
4464
4465 // Add the instructions that need to be reanalyzed to the worklist, and
4466 // reset the changed indicator.
4467 Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4468 InstsToReanalyze.clear();
4469 Changed = false;
4470
4471 while (!Worklist.empty()) {
4472 auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4473
4474 // We can't sink an instruction if it is a phi node, is already in the
4475 // predicated block, is not in the loop, or may have side effects.
4476 if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4477 !VectorLoop->contains(I) || I->mayHaveSideEffects())
4478 continue;
4479
4480 // It's legal to sink the instruction if all its uses occur in the
4481 // predicated block. Otherwise, there's nothing to do yet, and we may
4482 // need to reanalyze the instruction.
4483 if (!all_of(I->uses(), isBlockOfUsePredicated)) {
4484 InstsToReanalyze.push_back(I);
4485 continue;
4486 }
4487
4488 // Move the instruction to the beginning of the predicated block, and add
4489 // it's operands to the worklist.
4490 I->moveBefore(&*PredBB->getFirstInsertionPt());
4491 Worklist.insert(I->op_begin(), I->op_end());
4492
4493 // The sinking may have enabled other instructions to be sunk, so we will
4494 // need to iterate.
4495 Changed = true;
4496 }
4497 } while (Changed);
4498}
4499
4500void InnerLoopVectorizer::predicateInstructions() {
4501
4502 // For each instruction I marked for predication on value C, split I into its
4503 // own basic block to form an if-then construct over C. Since I may be fed by
4504 // an extractelement instruction or other scalar operand, we try to
4505 // iteratively sink its scalar operands into the predicated block. If I feeds
4506 // an insertelement instruction, we try to move this instruction into the
4507 // predicated block as well. For non-void types, a phi node will be created
4508 // for the resulting value (either vector or scalar).
4509 //
4510 // So for some predicated instruction, e.g. the conditional sdiv in:
4511 //
4512 // for.body:
4513 // ...
4514 // %add = add nsw i32 %mul, %0
4515 // %cmp5 = icmp sgt i32 %2, 7
4516 // br i1 %cmp5, label %if.then, label %if.end
4517 //
4518 // if.then:
4519 // %div = sdiv i32 %0, %1
4520 // br label %if.end
4521 //
4522 // if.end:
4523 // %x.0 = phi i32 [ %div, %if.then ], [ %add, %for.body ]
4524 //
4525 // the sdiv at this point is scalarized and if-converted using a select.
4526 // The inactive elements in the vector are not used, but the predicated
4527 // instruction is still executed for all vector elements, essentially:
4528 //
4529 // vector.body:
4530 // ...
4531 // %17 = add nsw <2 x i32> %16, %wide.load
4532 // %29 = extractelement <2 x i32> %wide.load, i32 0
4533 // %30 = extractelement <2 x i32> %wide.load51, i32 0
4534 // %31 = sdiv i32 %29, %30
4535 // %32 = insertelement <2 x i32> undef, i32 %31, i32 0
4536 // %35 = extractelement <2 x i32> %wide.load, i32 1
4537 // %36 = extractelement <2 x i32> %wide.load51, i32 1
4538 // %37 = sdiv i32 %35, %36
4539 // %38 = insertelement <2 x i32> %32, i32 %37, i32 1
4540 // %predphi = select <2 x i1> %26, <2 x i32> %38, <2 x i32> %17
4541 //
4542 // Predication will now re-introduce the original control flow to avoid false
4543 // side-effects by the sdiv instructions on the inactive elements, yielding
4544 // (after cleanup):
4545 //
4546 // vector.body:
4547 // ...
4548 // %5 = add nsw <2 x i32> %4, %wide.load
4549 // %8 = icmp sgt <2 x i32> %wide.load52, <i32 7, i32 7>
4550 // %9 = extractelement <2 x i1> %8, i32 0
4551 // br i1 %9, label %pred.sdiv.if, label %pred.sdiv.continue
4552 //
4553 // pred.sdiv.if:
4554 // %10 = extractelement <2 x i32> %wide.load, i32 0
4555 // %11 = extractelement <2 x i32> %wide.load51, i32 0
4556 // %12 = sdiv i32 %10, %11
4557 // %13 = insertelement <2 x i32> undef, i32 %12, i32 0
4558 // br label %pred.sdiv.continue
4559 //
4560 // pred.sdiv.continue:
4561 // %14 = phi <2 x i32> [ undef, %vector.body ], [ %13, %pred.sdiv.if ]
4562 // %15 = extractelement <2 x i1> %8, i32 1
4563 // br i1 %15, label %pred.sdiv.if54, label %pred.sdiv.continue55
4564 //
4565 // pred.sdiv.if54:
4566 // %16 = extractelement <2 x i32> %wide.load, i32 1
4567 // %17 = extractelement <2 x i32> %wide.load51, i32 1
4568 // %18 = sdiv i32 %16, %17
4569 // %19 = insertelement <2 x i32> %14, i32 %18, i32 1
4570 // br label %pred.sdiv.continue55
4571 //
4572 // pred.sdiv.continue55:
4573 // %20 = phi <2 x i32> [ %14, %pred.sdiv.continue ], [ %19, %pred.sdiv.if54 ]
4574 // %predphi = select <2 x i1> %8, <2 x i32> %20, <2 x i32> %5
4575
4576 for (auto KV : PredicatedInstructions) {
4577 BasicBlock::iterator I(KV.first);
4578 BasicBlock *Head = I->getParent();
4579 auto *BB = SplitBlock(Head, &*std::next(I), DT, LI);
4580 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
4581 /*BranchWeights=*/nullptr, DT, LI);
4582 I->moveBefore(T);
4583 sinkScalarOperands(&*I);
4584
4585 I->getParent()->setName(Twine("pred.") + I->getOpcodeName() + ".if");
4586 BB->setName(Twine("pred.") + I->getOpcodeName() + ".continue");
4587
4588 // If the instruction is non-void create a Phi node at reconvergence point.
4589 if (!I->getType()->isVoidTy()) {
4590 Value *IncomingTrue = nullptr;
4591 Value *IncomingFalse = nullptr;
4592
4593 if (I->hasOneUse() && isa<InsertElementInst>(*I->user_begin())) {
4594 // If the predicated instruction is feeding an insert-element, move it
4595 // into the Then block; Phi node will be created for the vector.
4596 InsertElementInst *IEI = cast<InsertElementInst>(*I->user_begin());
4597 IEI->moveBefore(T);
4598 IncomingTrue = IEI; // the new vector with the inserted element.
4599 IncomingFalse = IEI->getOperand(0); // the unmodified vector
4600 } else {
4601 // Phi node will be created for the scalar predicated instruction.
4602 IncomingTrue = &*I;
4603 IncomingFalse = UndefValue::get(I->getType());
4604 }
4605
4606 BasicBlock *PostDom = I->getParent()->getSingleSuccessor();
4607 assert(PostDom && "Then block has multiple successors")((PostDom && "Then block has multiple successors") ? static_cast
<void> (0) : __assert_fail ("PostDom && \"Then block has multiple successors\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4607, __PRETTY_FUNCTION__))
;
4608 PHINode *Phi =
4609 PHINode::Create(IncomingTrue->getType(), 2, "", &PostDom->front());
4610 IncomingTrue->replaceAllUsesWith(Phi);
4611 Phi->addIncoming(IncomingFalse, Head);
4612 Phi->addIncoming(IncomingTrue, I->getParent());
4613 }
4614 }
4615
4616 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (false
)
;
4617}
4618
4619InnerLoopVectorizer::VectorParts
4620InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
4621 assert(is_contained(predecessors(Dst), Src) && "Invalid edge")((is_contained(predecessors(Dst), Src) && "Invalid edge"
) ? static_cast<void> (0) : __assert_fail ("is_contained(predecessors(Dst), Src) && \"Invalid edge\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4621, __PRETTY_FUNCTION__))
;
4622
4623 // Look for cached value.
4624 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
4625 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
4626 if (ECEntryIt != MaskCache.end())
4627 return ECEntryIt->second;
4628
4629 VectorParts SrcMask = createBlockInMask(Src);
4630
4631 // The terminator has to be a branch inst!
4632 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
4633 assert(BI && "Unexpected terminator found")((BI && "Unexpected terminator found") ? static_cast<
void> (0) : __assert_fail ("BI && \"Unexpected terminator found\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4633, __PRETTY_FUNCTION__))
;
4634
4635 if (BI->isConditional()) {
4636 VectorParts EdgeMask = getVectorValue(BI->getCondition());
4637
4638 if (BI->getSuccessor(0) != Dst)
4639 for (unsigned part = 0; part < UF; ++part)
4640 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4641
4642 for (unsigned part = 0; part < UF; ++part)
4643 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4644
4645 MaskCache[Edge] = EdgeMask;
4646 return EdgeMask;
4647 }
4648
4649 MaskCache[Edge] = SrcMask;
4650 return SrcMask;
4651}
4652
4653InnerLoopVectorizer::VectorParts
4654InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4655 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4655, __PRETTY_FUNCTION__))
;
4656
4657 // Loop incoming mask is all-one.
4658 if (OrigLoop->getHeader() == BB) {
4659 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4660 return getVectorValue(C);
4661 }
4662
4663 // This is the block mask. We OR all incoming edges, and with zero.
4664 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4665 VectorParts BlockMask = getVectorValue(Zero);
4666
4667 // For each pred:
4668 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4669 VectorParts EM = createEdgeMask(*it, BB);
4670 for (unsigned part = 0; part < UF; ++part)
4671 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4672 }
4673
4674 return BlockMask;
4675}
4676
4677void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
4678 unsigned VF) {
4679 PHINode *P = cast<PHINode>(PN);
4680 // In order to support recurrences we need to be able to vectorize Phi nodes.
4681 // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4682 // stage #1: We create a new vector PHI node with no incoming edges. We'll use
4683 // this value when we vectorize all of the instructions that use the PHI.
4684 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4685 VectorParts Entry(UF);
4686 for (unsigned part = 0; part < UF; ++part) {
4687 // This is phase one of vectorizing PHIs.
4688 Type *VecTy =
4689 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4690 Entry[part] = PHINode::Create(
4691 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4692 }
4693 VectorLoopValueMap.initVector(P, Entry);
4694 return;
4695 }
4696
4697 setDebugLocFromInst(Builder, P);
4698 // Check for PHI nodes that are lowered to vector selects.
4699 if (P->getParent() != OrigLoop->getHeader()) {
4700 // We know that all PHIs in non-header blocks are converted into
4701 // selects, so we don't have to worry about the insertion order and we
4702 // can just use the builder.
4703 // At this point we generate the predication tree. There may be
4704 // duplications since this is a simple recursive scan, but future
4705 // optimizations will clean it up.
4706
4707 unsigned NumIncoming = P->getNumIncomingValues();
4708
4709 // Generate a sequence of selects of the form:
4710 // SELECT(Mask3, In3,
4711 // SELECT(Mask2, In2,
4712 // ( ...)))
4713 VectorParts Entry(UF);
4714 for (unsigned In = 0; In < NumIncoming; In++) {
4715 VectorParts Cond =
4716 createEdgeMask(P->getIncomingBlock(In), P->getParent());
4717 const VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
4718
4719 for (unsigned part = 0; part < UF; ++part) {
4720 // We might have single edge PHIs (blocks) - use an identity
4721 // 'select' for the first PHI operand.
4722 if (In == 0)
4723 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In0[part]);
4724 else
4725 // Select between the current value and the previous incoming edge
4726 // based on the incoming mask.
4727 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], Entry[part],
4728 "predphi");
4729 }
4730 }
4731 VectorLoopValueMap.initVector(P, Entry);
4732 return;
4733 }
4734
4735 // This PHINode must be an induction variable.
4736 // Make sure that we know about it.
4737 assert(Legal->getInductionVars()->count(P) && "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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4737, __PRETTY_FUNCTION__))
;
4738
4739 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4740 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4741
4742 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4743 // which can be found from the original scalar operations.
4744 switch (II.getKind()) {
4745 case InductionDescriptor::IK_NoInduction:
4746 llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4746)
;
4747 case InductionDescriptor::IK_IntInduction:
4748 case InductionDescriptor::IK_FpInduction:
4749 return widenIntOrFpInduction(P);
4750 case InductionDescriptor::IK_PtrInduction: {
4751 // Handle the pointer induction variable case.
4752 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4752, __PRETTY_FUNCTION__))
;
4753 // This is the normalized GEP that starts counting at zero.
4754 Value *PtrInd = Induction;
4755 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4756 // Determine the number of scalars we need to generate for each unroll
4757 // iteration. If the instruction is uniform, we only need to generate the
4758 // first lane. Otherwise, we generate all VF values.
4759 unsigned Lanes = Cost->isUniformAfterVectorization(P, VF) ? 1 : VF;
4760 // These are the scalar results. Notice that we don't generate vector GEPs
4761 // because scalar GEPs result in better code.
4762 ScalarParts Entry(UF);
4763 for (unsigned Part = 0; Part < UF; ++Part) {
4764 Entry[Part].resize(VF);
4765 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4766 Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4767 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4768 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4769 SclrGep->setName("next.gep");
4770 Entry[Part][Lane] = SclrGep;
4771 }
4772 }
4773 VectorLoopValueMap.initScalar(P, Entry);
4774 return;
4775 }
4776 }
4777}
4778
4779/// A helper function for checking whether an integer division-related
4780/// instruction may divide by zero (in which case it must be predicated if
4781/// executed conditionally in the scalar code).
4782/// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4783/// Non-zero divisors that are non compile-time constants will not be
4784/// converted into multiplication, so we will still end up scalarizing
4785/// the division, but can do so w/o predication.
4786static bool mayDivideByZero(Instruction &I) {
4787 assert((I.getOpcode() == Instruction::UDiv ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction
::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode()
== Instruction::SRem) && "Unexpected instruction") ?
static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4791, __PRETTY_FUNCTION__))
4788 I.getOpcode() == Instruction::SDiv ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction
::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode()
== Instruction::SRem) && "Unexpected instruction") ?
static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4791, __PRETTY_FUNCTION__))
4789 I.getOpcode() == Instruction::URem ||(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction
::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode()
== Instruction::SRem) && "Unexpected instruction") ?
static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4791, __PRETTY_FUNCTION__))
4790 I.getOpcode() == Instruction::SRem) &&(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction
::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode()
== Instruction::SRem) && "Unexpected instruction") ?
static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4791, __PRETTY_FUNCTION__))
4791 "Unexpected instruction")(((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction
::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode()
== Instruction::SRem) && "Unexpected instruction") ?
static_cast<void> (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4791, __PRETTY_FUNCTION__))
;
4792 Value *Divisor = I.getOperand(1);
4793 auto *CInt = dyn_cast<ConstantInt>(Divisor);
4794 return !CInt || CInt->isZero();
4795}
4796
4797void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB) {
4798 // For each instruction in the old loop.
4799 for (Instruction &I : *BB) {
4800
4801 // If the instruction will become trivially dead when vectorized, we don't
4802 // need to generate it.
4803 if (DeadInstructions.count(&I))
4804 continue;
4805
4806 // Scalarize instructions that should remain scalar after vectorization.
4807 if (VF > 1 &&
4808 !(isa<BranchInst>(&I) || isa<PHINode>(&I) ||
4809 isa<DbgInfoIntrinsic>(&I)) &&
4810 shouldScalarizeInstruction(&I)) {
4811 scalarizeInstruction(&I, Legal->isScalarWithPredication(&I));
4812 continue;
4813 }
4814
4815 switch (I.getOpcode()) {
4816 case Instruction::Br:
4817 // Nothing to do for PHIs and BR, since we already took care of the
4818 // loop control flow instructions.
4819 continue;
4820 case Instruction::PHI: {
4821 // Vectorize PHINodes.
4822 widenPHIInstruction(&I, UF, VF);
4823 continue;
4824 } // End of PHI.
4825
4826 case Instruction::UDiv:
4827 case Instruction::SDiv:
4828 case Instruction::SRem:
4829 case Instruction::URem:
4830 // Scalarize with predication if this instruction may divide by zero and
4831 // block execution is conditional, otherwise fallthrough.
4832 if (Legal->isScalarWithPredication(&I)) {
4833 scalarizeInstruction(&I, true);
4834 continue;
4835 }
4836 case Instruction::Add:
4837 case Instruction::FAdd:
4838 case Instruction::Sub:
4839 case Instruction::FSub:
4840 case Instruction::Mul:
4841 case Instruction::FMul:
4842 case Instruction::FDiv:
4843 case Instruction::FRem:
4844 case Instruction::Shl:
4845 case Instruction::LShr:
4846 case Instruction::AShr:
4847 case Instruction::And:
4848 case Instruction::Or:
4849 case Instruction::Xor: {
4850 // Just widen binops.
4851 auto *BinOp = cast<BinaryOperator>(&I);
4852 setDebugLocFromInst(Builder, BinOp);
4853 const VectorParts &A = getVectorValue(BinOp->getOperand(0));
4854 const VectorParts &B = getVectorValue(BinOp->getOperand(1));
4855
4856 // Use this vector value for all users of the original instruction.
4857 VectorParts Entry(UF);
4858 for (unsigned Part = 0; Part < UF; ++Part) {
4859 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4860
4861 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4862 VecOp->copyIRFlags(BinOp);
4863
4864 Entry[Part] = V;
4865 }
4866
4867 VectorLoopValueMap.initVector(&I, Entry);
4868 addMetadata(Entry, BinOp);
4869 break;
4870 }
4871 case Instruction::Select: {
4872 // Widen selects.
4873 // If the selector is loop invariant we can create a select
4874 // instruction with a scalar condition. Otherwise, use vector-select.
4875 auto *SE = PSE.getSE();
4876 bool InvariantCond =
4877 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4878 setDebugLocFromInst(Builder, &I);
4879
4880 // The condition can be loop invariant but still defined inside the
4881 // loop. This means that we can't just use the original 'cond' value.
4882 // We have to take the 'vectorized' value and pick the first lane.
4883 // Instcombine will make this a no-op.
4884 const VectorParts &Cond = getVectorValue(I.getOperand(0));
4885 const VectorParts &Op0 = getVectorValue(I.getOperand(1));
4886 const VectorParts &Op1 = getVectorValue(I.getOperand(2));
4887
4888 auto *ScalarCond = getScalarValue(I.getOperand(0), 0, 0);
4889
4890 VectorParts Entry(UF);
4891 for (unsigned Part = 0; Part < UF; ++Part) {
4892 Entry[Part] = Builder.CreateSelect(
4893 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4894 }
4895
4896 VectorLoopValueMap.initVector(&I, Entry);
4897 addMetadata(Entry, &I);
4898 break;
4899 }
4900
4901 case Instruction::ICmp:
4902 case Instruction::FCmp: {
4903 // Widen compares. Generate vector compares.
4904 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4905 auto *Cmp = dyn_cast<CmpInst>(&I);
4906 setDebugLocFromInst(Builder, Cmp);
4907 const VectorParts &A = getVectorValue(Cmp->getOperand(0));
4908 const VectorParts &B = getVectorValue(Cmp->getOperand(1));
4909 VectorParts Entry(UF);
4910 for (unsigned Part = 0; Part < UF; ++Part) {
4911 Value *C = nullptr;
4912 if (FCmp) {
4913 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4914 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4915 } else {
4916 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4917 }
4918 Entry[Part] = C;
4919 }
4920
4921 VectorLoopValueMap.initVector(&I, Entry);
4922 addMetadata(Entry, &I);
4923 break;
4924 }
4925
4926 case Instruction::Store:
4927 case Instruction::Load:
4928 vectorizeMemoryInstruction(&I);
4929 break;
4930 case Instruction::ZExt:
4931 case Instruction::SExt:
4932 case Instruction::FPToUI:
4933 case Instruction::FPToSI:
4934 case Instruction::FPExt:
4935 case Instruction::PtrToInt:
4936 case Instruction::IntToPtr:
4937 case Instruction::SIToFP:
4938 case Instruction::UIToFP:
4939 case Instruction::Trunc:
4940 case Instruction::FPTrunc:
4941 case Instruction::BitCast: {
4942 auto *CI = dyn_cast<CastInst>(&I);
4943 setDebugLocFromInst(Builder, CI);
4944
4945 // Optimize the special case where the source is a constant integer
4946 // induction variable. Notice that we can only optimize the 'trunc' case
4947 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4948 // (c) other casts depend on pointer size.
4949 if (Cost->isOptimizableIVTruncate(CI, VF)) {
4950 widenIntOrFpInduction(cast<PHINode>(CI->getOperand(0)),
4951 cast<TruncInst>(CI));
4952 break;
4953 }
4954
4955 /// Vectorize casts.
4956 Type *DestTy =
4957 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4958
4959 const VectorParts &A = getVectorValue(CI->getOperand(0));
4960 VectorParts Entry(UF);
4961 for (unsigned Part = 0; Part < UF; ++Part)
4962 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4963 VectorLoopValueMap.initVector(&I, Entry);
4964 addMetadata(Entry, &I);
4965 break;
4966 }
4967
4968 case Instruction::Call: {
4969 // Ignore dbg intrinsics.
4970 if (isa<DbgInfoIntrinsic>(I))
4971 break;
4972 setDebugLocFromInst(Builder, &I);
4973
4974 Module *M = BB->getParent()->getParent();
4975 auto *CI = cast<CallInst>(&I);
4976
4977 StringRef FnName = CI->getCalledFunction()->getName();
4978 Function *F = CI->getCalledFunction();
4979 Type *RetTy = ToVectorTy(CI->getType(), VF);
4980 SmallVector<Type *, 4> Tys;
4981 for (Value *ArgOperand : CI->arg_operands())
4982 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4983
4984 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4985 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4986 ID == Intrinsic::lifetime_start)) {
4987 scalarizeInstruction(&I);
4988 break;
4989 }
4990 // The flag shows whether we use Intrinsic or a usual Call for vectorized
4991 // version of the instruction.
4992 // Is it beneficial to perform intrinsic call compared to lib call?
4993 bool NeedToScalarize;
4994 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4995 bool UseVectorIntrinsic =
4996 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4997 if (!UseVectorIntrinsic && NeedToScalarize) {
4998 scalarizeInstruction(&I);
4999 break;
5000 }
5001
5002 VectorParts Entry(UF);
5003 for (unsigned Part = 0; Part < UF; ++Part) {
5004 SmallVector<Value *, 4> Args;
5005 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
5006 Value *Arg = CI->getArgOperand(i);
5007 // Some intrinsics have a scalar argument - don't replace it with a
5008 // vector.
5009 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
5010 const VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
5011 Arg = VectorArg[Part];
5012 }
5013 Args.push_back(Arg);
5014 }
5015
5016 Function *VectorF;
5017 if (UseVectorIntrinsic) {
5018 // Use vector version of the intrinsic.
5019 Type *TysForDecl[] = {CI->getType()};
5020 if (VF > 1)
5021 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
5022 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
5023 } else {
5024 // Use vector version of the library call.
5025 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
5026 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5026, __PRETTY_FUNCTION__))
;
5027 VectorF = M->getFunction(VFnName);
5028 if (!VectorF) {
5029 // Generate a declaration
5030 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
5031 VectorF =
5032 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
5033 VectorF->copyAttributesFrom(F);
5034 }
5035 }
5036 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5036, __PRETTY_FUNCTION__))
;
5037
5038 SmallVector<OperandBundleDef, 1> OpBundles;
5039 CI->getOperandBundlesAsDefs(OpBundles);
5040 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
5041
5042 if (isa<FPMathOperator>(V))
5043 V->copyFastMathFlags(CI);
5044
5045 Entry[Part] = V;
5046 }
5047
5048 VectorLoopValueMap.initVector(&I, Entry);
5049 addMetadata(Entry, &I);
5050 break;
5051 }
5052
5053 default:
5054 // All other instructions are unsupported. Scalarize them.
5055 scalarizeInstruction(&I);
5056 break;
5057 } // end of switch.
5058 } // end of for_each instr.
5059}
5060
5061void InnerLoopVectorizer::updateAnalysis() {
5062 // Forget the original basic block.
5063 PSE.getSE()->forgetLoop(OrigLoop);
5064
5065 // Update the dominator tree information.
5066 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5067, __PRETTY_FUNCTION__))
5067 "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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5067, __PRETTY_FUNCTION__))
;
5068
5069 // We don't predicate stores by this point, so the vector body should be a
5070 // single loop.
5071 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
5072
5073 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody);
5074 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
5075 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
5076 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
5077
5078 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (false
)
;
5079}
5080
5081/// \brief Check whether it is safe to if-convert this phi node.
5082///
5083/// Phi nodes with constant expressions that can trap are not safe to if
5084/// convert.
5085static bool canIfConvertPHINodes(BasicBlock *BB) {
5086 for (Instruction &I : *BB) {
5087 auto *Phi = dyn_cast<PHINode>(&I);
5088 if (!Phi)
5089 return true;
5090 for (Value *V : Phi->incoming_values())
5091 if (auto *C = dyn_cast<Constant>(V))
5092 if (C->canTrap())
5093 return false;
5094 }
5095 return true;
5096}
5097
5098bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
5099 if (!EnableIfConversion) {
5100 ORE->emit(createMissedAnalysis("IfConversionDisabled")
5101 << "if-conversion is disabled");
5102 return false;
5103 }
5104
5105 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5105, __PRETTY_FUNCTION__))
;
5106
5107 // A list of pointers that we can safely read and write to.
5108 SmallPtrSet<Value *, 8> SafePointes;
5109
5110 // Collect safe addresses.
5111 for (BasicBlock *BB : TheLoop->blocks()) {
5112 if (blockNeedsPredication(BB))
5113 continue;
5114
5115 for (Instruction &I : *BB)
5116 if (auto *Ptr = getPointerOperand(&I))
5117 SafePointes.insert(Ptr);
5118 }
5119
5120 // Collect the blocks that need predication.
5121 BasicBlock *Header = TheLoop->getHeader();
5122 for (BasicBlock *BB : TheLoop->blocks()) {
5123 // We don't support switch statements inside loops.
5124 if (!isa<BranchInst>(BB->getTerminator())) {
5125 ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5126 << "loop contains a switch statement");
5127 return false;
5128 }
5129
5130 // We must be able to predicate all blocks that need to be predicated.
5131 if (blockNeedsPredication(BB)) {
5132 if (!blockCanBePredicated(BB, SafePointes)) {
5133 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5134 << "control flow cannot be substituted for a select");
5135 return false;
5136 }
5137 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5138 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5139 << "control flow cannot be substituted for a select");
5140 return false;
5141 }
5142 }
5143
5144 // We can if-convert this loop.
5145 return true;
5146}
5147
5148bool LoopVectorizationLegality::canVectorize() {
5149 // We must have a loop in canonical form. Loops with indirectbr in them cannot
5150 // be canonicalized.
5151 if (!TheLoop->getLoopPreheader()) {
5152 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5153 << "loop control flow is not understood by vectorizer");
5154 return false;
5155 }
5156
5157 // FIXME: The code is currently dead, since the loop gets sent to
5158 // LoopVectorizationLegality is already an innermost loop.
5159 //
5160 // We can only vectorize innermost loops.
5161 if (!TheLoop->empty()) {
5162 ORE->emit(createMissedAnalysis("NotInnermostLoop")
5163 << "loop is not the innermost loop");
5164 return false;
5165 }
5166
5167 // We must have a single backedge.
5168 if (TheLoop->getNumBackEdges() != 1) {
5169 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5170 << "loop control flow is not understood by vectorizer");
5171 return false;
5172 }
5173
5174 // We must have a single exiting block.
5175 if (!TheLoop->getExitingBlock()) {
5176 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5177 << "loop control flow is not understood by vectorizer");
5178 return false;
5179 }
5180
5181 // We only handle bottom-tested loops, i.e. loop in which the condition is
5182 // checked at the end of each iteration. With that we can assume that all
5183 // instructions in the loop are executed the same number of times.
5184 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5185 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5186 << "loop control flow is not understood by vectorizer");
5187 return false;
5188 }
5189
5190 // We need to have a loop header.
5191 DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop: " <<
TheLoop->getHeader()->getName() << '\n'; } } while
(false)
5192 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop: " <<
TheLoop->getHeader()->getName() << '\n'; } } while
(false)
;
5193
5194 // Check if we can if-convert non-single-bb loops.
5195 unsigned NumBlocks = TheLoop->getNumBlocks();
5196 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5197 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 (false)
;
5198 return false;
5199 }
5200
5201 // ScalarEvolution needs to be able to find the exit count.
5202 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
5203 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
5204 ORE->emit(createMissedAnalysis("CantComputeNumberOfIterations")
5205 << "could not determine number of loop iterations");
5206 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 (false)
;
5207 return false;
5208 }
5209
5210 // Check if we can vectorize the instructions and CFG in this loop.
5211 if (!canVectorizeInstrs()) {
5212 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 (false)
;
5213 return false;
5214 }
5215
5216 // Go over each instruction and look at memory deps.
5217 if (!canVectorizeMemory()) {
5218 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 (false)
;
5219 return false;
5220 }
5221
5222 DEBUG(dbgs() << "LV: We can vectorize this loop"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerChecking()->Need ? " (with a runtime bound check)"
: "") << "!\n"; } } while (false)
5223 << (LAI->getRuntimePointerChecking()->Needdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerChecking()->Need ? " (with a runtime bound check)"
: "") << "!\n"; } } while (false)
5224 ? " (with a runtime bound check)"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerChecking()->Need ? " (with a runtime bound check)"
: "") << "!\n"; } } while (false)
5225 : "")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerChecking()->Need ? " (with a runtime bound check)"
: "") << "!\n"; } } while (false)
5226 << "!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: We can vectorize this loop"
<< (LAI->getRuntimePointerChecking()->Need ? " (with a runtime bound check)"
: "") << "!\n"; } } while (false)
;
5227
5228 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5229
5230 // If an override option has been passed in for interleaved accesses, use it.
5231 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5232 UseInterleaved = EnableInterleavedMemAccesses;
5233
5234 // Analyze interleaved memory accesses.
5235 if (UseInterleaved)
5236 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5237
5238 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5239 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5240 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5241
5242 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5243 ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5244 << "Too many SCEV assumptions need to be made and checked "
5245 << "at runtime");
5246 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Too many SCEV checks needed.\n"
; } } while (false)
;
5247 return false;
5248 }
5249
5250 // Okay! We can vectorize. At this point we don't have any other mem analysis
5251 // which may limit our maximum vectorization factor, so just return true with
5252 // no restrictions.
5253 return true;
5254}
5255
5256static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
5257 if (Ty->isPointerTy())
5258 return DL.getIntPtrType(Ty);
5259
5260 // It is possible that char's or short's overflow when we ask for the loop's
5261 // trip count, work around this by changing the type size.
5262 if (Ty->getScalarSizeInBits() < 32)
5263 return Type::getInt32Ty(Ty->getContext());
5264
5265 return Ty;
5266}
5267
5268static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
5269 Ty0 = convertPointerToIntegerType(DL, Ty0);
5270 Ty1 = convertPointerToIntegerType(DL, Ty1);
5271 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
5272 return Ty0;
5273 return Ty1;
5274}
5275
5276/// \brief Check that the instruction has outside loop users and is not an
5277/// identified reduction variable.
5278static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
5279 SmallPtrSetImpl<Value *> &AllowedExit) {
5280 // Reduction and Induction instructions are allowed to have exit users. All
5281 // other instructions must not have external users.
5282 if (!AllowedExit.count(Inst))
5283 // Check that all of the users of the loop are inside the BB.
5284 for (User *U : Inst->users()) {
5285 Instruction *UI = cast<Instruction>(U);
5286 // This user may be a reduction exit value.
5287 if (!TheLoop->contains(UI)) {
5288 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 (false)
;
5289 return true;
5290 }
5291 }
5292 return false;
5293}
5294
5295void LoopVectorizationLegality::addInductionPhi(
5296 PHINode *Phi, const InductionDescriptor &ID,
5297 SmallPtrSetImpl<Value *> &AllowedExit) {
5298 Inductions[Phi] = ID;
5299 Type *PhiTy = Phi->getType();
5300 const DataLayout &DL = Phi->getModule()->getDataLayout();
5301
5302 // Get the widest type.
5303 if (!PhiTy->isFloatingPointTy()) {
5304 if (!WidestIndTy)
5305 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
5306 else
5307 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
5308 }
5309
5310 // Int inductions are special because we only allow one IV.
5311 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
5312 ID.getConstIntStepValue() &&
5313 ID.getConstIntStepValue()->isOne() &&
5314 isa<Constant>(ID.getStartValue()) &&
5315 cast<Constant>(ID.getStartValue())->isNullValue()) {
5316
5317 // Use the phi node with the widest type as induction. Use the last
5318 // one if there are multiple (no good reason for doing this other
5319 // than it is expedient). We've checked that it begins at zero and
5320 // steps by one, so this is a canonical induction variable.
5321 if (!PrimaryInduction || PhiTy == WidestIndTy)
5322 PrimaryInduction = Phi;
5323 }
5324
5325 // Both the PHI node itself, and the "post-increment" value feeding
5326 // back into the PHI node may have external users.
5327 AllowedExit.insert(Phi);
5328 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
5329
5330 DEBUG(dbgs() << "LV: Found an induction variable.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an induction variable.\n"
; } } while (false)
;
5331 return;
5332}
5333
5334bool LoopVectorizationLegality::canVectorizeInstrs() {
5335 BasicBlock *Header = TheLoop->getHeader();
5336
5337 // Look for the attribute signaling the absence of NaNs.
5338 Function &F = *Header->getParent();
5339 HasFunNoNaNAttr =
5340 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
5341
5342 // For each block in the loop.
5343 for (BasicBlock *BB : TheLoop->blocks()) {
5344 // Scan the instructions in the block and look for hazards.
5345 for (Instruction &I : *BB) {
5346 if (auto *Phi = dyn_cast<PHINode>(&I)) {
5347 Type *PhiTy = Phi->getType();
5348 // Check that this PHI type is allowed.
5349 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
5350 !PhiTy->isPointerTy()) {
5351 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5352 << "loop control flow is not understood by vectorizer");
5353 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 (false)
;
5354 return false;
5355 }
5356
5357 // If this PHINode is not in the header block, then we know that we
5358 // can convert it to select during if-conversion. No need to check if
5359 // the PHIs in this block are induction or reduction variables.
5360 if (BB != Header) {
5361 // Check that this instruction has no outside users or is an
5362 // identified reduction value with an outside user.
5363 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
5364 continue;
5365 ORE->emit(createMissedAnalysis("NeitherInductionNorReduction", Phi)
5366 << "value could not be identified as "
5367 "an induction or reduction variable");
5368 return false;
5369 }
5370
5371 // We only allow if-converted PHIs with exactly two incoming values.
5372 if (Phi->getNumIncomingValues() != 2) {
5373 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5374 << "control flow not understood by vectorizer");
5375 DEBUG(dbgs() << "LV: Found an invalid PHI.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an invalid PHI.\n"
; } } while (false)
;
5376 return false;
5377 }
5378
5379 RecurrenceDescriptor RedDes;
5380 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
5381 if (RedDes.hasUnsafeAlgebra())
5382 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
5383 AllowedExit.insert(RedDes.getLoopExitInstr());
5384 Reductions[Phi] = RedDes;
5385 continue;
5386 }
5387
5388 InductionDescriptor ID;
5389 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
5390 addInductionPhi(Phi, ID, AllowedExit);
5391 if (ID.hasUnsafeAlgebra() && !HasFunNoNaNAttr)
5392 Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
5393 continue;
5394 }
5395
5396 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
5397 FirstOrderRecurrences.insert(Phi);
5398 continue;
5399 }
5400
5401 // As a last resort, coerce the PHI to a AddRec expression
5402 // and re-try classifying it a an induction PHI.
5403 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
5404 addInductionPhi(Phi, ID, AllowedExit);
5405 continue;
5406 }
5407
5408 ORE->emit(createMissedAnalysis("NonReductionValueUsedOutsideLoop", Phi)
5409 << "value that could not be identified as "
5410 "reduction is used outside the loop");
5411 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 (false)
;
5412 return false;
5413 } // end of PHI handling
5414
5415 // We handle calls that:
5416 // * Are debug info intrinsics.
5417 // * Have a mapping to an IR intrinsic.
5418 // * Have a vector version available.
5419 auto *CI = dyn_cast<CallInst>(&I);
5420 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
5421 !isa<DbgInfoIntrinsic>(CI) &&
5422 !(CI->getCalledFunction() && TLI &&
5423 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
5424 ORE->emit(createMissedAnalysis("CantVectorizeCall", CI)
5425 << "call instruction cannot be vectorized");
5426 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 (false)
;
5427 return false;
5428 }
5429
5430 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
5431 // second argument is the same (i.e. loop invariant)
5432 if (CI && hasVectorInstrinsicScalarOpd(
5433 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
5434 auto *SE = PSE.getSE();
5435 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
5436 ORE->emit(createMissedAnalysis("CantVectorizeIntrinsic", CI)
5437 << "intrinsic instruction cannot be vectorized");
5438 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found unvectorizable intrinsic "
<< *CI << "\n"; } } while (false)
;
5439 return false;
5440 }
5441 }
5442
5443 // Check that the instruction return type is vectorizable.
5444 // Also, we can't vectorize extractelement instructions.
5445 if ((!VectorType::isValidElementType(I.getType()) &&
5446 !I.getType()->isVoidTy()) ||
5447 isa<ExtractElementInst>(I)) {
5448 ORE->emit(createMissedAnalysis("CantVectorizeInstructionReturnType", &I)
5449 << "instruction return type cannot be vectorized");
5450 DEBUG(dbgs() << "LV: Found unvectorizable type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found unvectorizable type.\n"
; } } while (false)
;
5451 return false;
5452 }
5453
5454 // Check that the stored type is vectorizable.
5455 if (auto *ST = dyn_cast<StoreInst>(&I)) {
5456 Type *T = ST->getValueOperand()->getType();
5457 if (!VectorType::isValidElementType(T)) {
5458 ORE->emit(createMissedAnalysis("CantVectorizeStore", ST)
5459 << "store instruction cannot be vectorized");
5460 return false;
5461 }
5462
5463 // FP instructions can allow unsafe algebra, thus vectorizable by
5464 // non-IEEE-754 compliant SIMD units.
5465 // This applies to floating-point math operations and calls, not memory
5466 // operations, shuffles, or casts, as they don't change precision or
5467 // semantics.
5468 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
5469 !I.hasUnsafeAlgebra()) {
5470 DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found FP op with unsafe algebra.\n"
; } } while (false)
;
5471 Hints->setPotentiallyUnsafe();
5472 }
5473
5474 // Reduction instructions are allowed to have exit users.
5475 // All other instructions must not have external users.
5476 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
5477 ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)
5478 << "value cannot be used outside the loop");
5479 return false;
5480 }
5481
5482 } // next instr.
5483 }
5484
5485 if (!PrimaryInduction) {
5486 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 (false)
;
5487 if (Inductions.empty()) {
5488 ORE->emit(createMissedAnalysis("NoInductionVariable")
5489 << "loop induction variable could not be identified");
5490 return false;
5491 }
5492 }
5493
5494 // Now we know the widest induction type, check if our found induction
5495 // is the same size. If it's not, unset it here and InnerLoopVectorizer
5496 // will create another.
5497 if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType())
5498 PrimaryInduction = nullptr;
5499
5500 return true;
5501}
5502
5503void LoopVectorizationCostModel::collectLoopScalars(unsigned VF) {
5504
5505 // We should not collect Scalars more than once per VF. Right now,
5506 // this function is called from collectUniformsAndScalars(), which
5507 // already does this check. Collecting Scalars for VF=1 does not make any
5508 // sense.
5509
5510 assert(VF >= 2 && !Scalars.count(VF) &&((VF >= 2 && !Scalars.count(VF) && "This function should not be visited twice for the same VF"
) ? static_cast<void> (0) : __assert_fail ("VF >= 2 && !Scalars.count(VF) && \"This function should not be visited twice for the same VF\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5511, __PRETTY_FUNCTION__))
5511 "This function should not be visited twice for the same VF")((VF >= 2 && !Scalars.count(VF) && "This function should not be visited twice for the same VF"
) ? static_cast<void> (0) : __assert_fail ("VF >= 2 && !Scalars.count(VF) && \"This function should not be visited twice for the same VF\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5511, __PRETTY_FUNCTION__))
;
5512
5513 // If an instruction is uniform after vectorization, it will remain scalar.
5514 Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end());
5515
5516 // Collect the getelementptr instructions that will not be vectorized. A
5517 // getelementptr instruction is only vectorized if it is used for a legal
5518 // gather or scatter operation.
5519 for (auto *BB : TheLoop->blocks())
5520 for (auto &I : *BB) {
5521 if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
5522 Scalars[VF].insert(GEP);
5523 continue;
5524 }
5525 auto *Ptr = getPointerOperand(&I);
5526 if (!Ptr)
5527 continue;
5528 auto *GEP = getGEPInstruction(Ptr);
5529 if (GEP && getWideningDecision(&I, VF) == CM_GatherScatter)
5530 Scalars[VF].erase(GEP);
5531 }
5532
5533 // An induction variable will remain scalar if all users of the induction
5534 // variable and induction variable update remain scalar.
5535 auto *Latch = TheLoop->getLoopLatch();
5536 for (auto &Induction : *Legal->getInductionVars()) {
5537 auto *Ind = Induction.first;
5538 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5539
5540 // Determine if all users of the induction variable are scalar after
5541 // vectorization.
5542 auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
5543 auto *I = cast<Instruction>(U);
5544 return I == IndUpdate || !TheLoop->contains(I) || Scalars[VF].count(I);
5545 });
5546 if (!ScalarInd)
5547 continue;
5548
5549 // Determine if all users of the induction variable update instruction are
5550 // scalar after vectorization.
5551 auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5552 auto *I = cast<Instruction>(U);
5553 return I == Ind || !TheLoop->contains(I) || Scalars[VF].count(I);
5554 });
5555 if (!ScalarIndUpdate)
5556 continue;
5557
5558 // The induction variable and its update instruction will remain scalar.
5559 Scalars[VF].insert(Ind);
5560 Scalars[VF].insert(IndUpdate);
5561 }
5562}
5563
5564bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
5565 if (!blockNeedsPredication(I->getParent()))
5566 return false;
5567 switch(I->getOpcode()) {
5568 default:
5569 break;
5570 case Instruction::Store:
5571 return !isMaskRequired(I);
5572 case Instruction::UDiv:
5573 case Instruction::SDiv:
5574 case Instruction::SRem:
5575 case Instruction::URem:
5576 return mayDivideByZero(*I);
5577 }
5578 return false;
5579}
5580
5581bool LoopVectorizationLegality::memoryInstructionCanBeWidened(Instruction *I,
5582 unsigned VF) {
5583 // Get and ensure we have a valid memory instruction.
5584 LoadInst *LI = dyn_cast<LoadInst>(I);
5585 StoreInst *SI = dyn_cast<StoreInst>(I);
5586 assert((LI || SI) && "Invalid memory instruction")(((LI || SI) && "Invalid memory instruction") ? static_cast
<void> (0) : __assert_fail ("(LI || SI) && \"Invalid memory instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5586, __PRETTY_FUNCTION__))
;
5587
5588 auto *Ptr = getPointerOperand(I);
5589
5590 // In order to be widened, the pointer should be consecutive, first of all.
5591 if (!isConsecutivePtr(Ptr))
5592 return false;
5593
5594 // If the instruction is a store located in a predicated block, it will be
5595 // scalarized.
5596 if (isScalarWithPredication(I))
5597 return false;
5598
5599 // If the instruction's allocated size doesn't equal it's type size, it
5600 // requires padding and will be scalarized.
5601 auto &DL = I->getModule()->getDataLayout();
5602 auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
5603 if (hasIrregularType(ScalarTy, DL, VF))
5604 return false;
5605
5606 return true;
5607}
5608
5609void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) {
5610
5611 // We should not collect Uniforms more than once per VF. Right now,
5612 // this function is called from collectUniformsAndScalars(), which
5613 // already does this check. Collecting Uniforms for VF=1 does not make any
5614 // sense.
5615
5616 assert(VF >= 2 && !Uniforms.count(VF) &&((VF >= 2 && !Uniforms.count(VF) && "This function should not be visited twice for the same VF"
) ? static_cast<void> (0) : __assert_fail ("VF >= 2 && !Uniforms.count(VF) && \"This function should not be visited twice for the same VF\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5617, __PRETTY_FUNCTION__))
5617 "This function should not be visited twice for the same VF")((VF >= 2 && !Uniforms.count(VF) && "This function should not be visited twice for the same VF"
) ? static_cast<void> (0) : __assert_fail ("VF >= 2 && !Uniforms.count(VF) && \"This function should not be visited twice for the same VF\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5617, __PRETTY_FUNCTION__))
;
5618
5619 // Visit the list of Uniforms. If we'll not find any uniform value, we'll
5620 // not analyze again. Uniforms.count(VF) will return 1.
5621 Uniforms[VF].clear();
5622
5623 // We now know that the loop is vectorizable!
5624 // Collect instructions inside the loop that will remain uniform after
5625 // vectorization.
5626
5627 // Global values, params and instructions outside of current loop are out of
5628 // scope.
5629 auto isOutOfScope = [&](Value *V) -> bool {
5630 Instruction *I = dyn_cast<Instruction>(V);
5631 return (!I || !TheLoop->contains(I));
5632 };
5633
5634 SetVector<Instruction *> Worklist;
5635 BasicBlock *Latch = TheLoop->getLoopLatch();
5636
5637 // Start with the conditional branch. If the branch condition is an
5638 // instruction contained in the loop that is only used by the branch, it is
5639 // uniform.
5640 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
5641 if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) {
5642 Worklist.insert(Cmp);
5643 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Cmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: "
<< *Cmp << "\n"; } } while (false)
;
5644 }
5645
5646 // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
5647 // are pointers that are treated like consecutive pointers during
5648 // vectorization. The pointer operands of interleaved accesses are an
5649 // example.
5650 SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
5651
5652 // Holds pointer operands of instructions that are possibly non-uniform.
5653 SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
5654
5655 auto isUniformDecision = [&](Instruction *I, unsigned VF) {
5656 InstWidening WideningDecision = getWideningDecision(I, VF);
5657 assert(WideningDecision != CM_Unknown &&((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment"
) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5658, __PRETTY_FUNCTION__))
5658 "Widening decision should be ready at this moment")((WideningDecision != CM_Unknown && "Widening decision should be ready at this moment"
) ? static_cast<void> (0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5658, __PRETTY_FUNCTION__))
;
5659
5660 return (WideningDecision == CM_Widen ||
5661 WideningDecision == CM_Interleave);
5662 };
5663 // Iterate over the instructions in the loop, and collect all
5664 // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
5665 // that a consecutive-like pointer operand will be scalarized, we collect it
5666 // in PossibleNonUniformPtrs instead. We use two sets here because a single
5667 // getelementptr instruction can be used by both vectorized and scalarized
5668 // memory instructions. For example, if a loop loads and stores from the same
5669 // location, but the store is conditional, the store will be scalarized, and
5670 // the getelementptr won't remain uniform.
5671 for (auto *BB : TheLoop->blocks())
5672 for (auto &I : *BB) {
5673
5674 // If there's no pointer operand, there's nothing to do.
5675 auto *Ptr = dyn_cast_or_null<Instruction>(getPointerOperand(&I));
5676 if (!Ptr)
5677 continue;
5678
5679 // True if all users of Ptr are memory accesses that have Ptr as their
5680 // pointer operand.
5681 auto UsersAreMemAccesses = all_of(Ptr->users(), [&](User *U) -> bool {
5682 return getPointerOperand(U) == Ptr;
5683 });
5684
5685 // Ensure the memory instruction will not be scalarized or used by
5686 // gather/scatter, making its pointer operand non-uniform. If the pointer
5687 // operand is used by any instruction other than a memory access, we
5688 // conservatively assume the pointer operand may be non-uniform.
5689 if (!UsersAreMemAccesses || !isUniformDecision(&I, VF))
5690 PossibleNonUniformPtrs.insert(Ptr);
5691
5692 // If the memory instruction will be vectorized and its pointer operand
5693 // is consecutive-like, or interleaving - the pointer operand should
5694 // remain uniform.
5695 else
5696 ConsecutiveLikePtrs.insert(Ptr);
5697 }
5698
5699 // Add to the Worklist all consecutive and consecutive-like pointers that
5700 // aren't also identified as possibly non-uniform.
5701 for (auto *V : ConsecutiveLikePtrs)
5702 if (!PossibleNonUniformPtrs.count(V)) {
5703 DEBUG(dbgs() << "LV: Found uniform instruction: " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: "
<< *V << "\n"; } } while (false)
;
5704 Worklist.insert(V);
5705 }
5706
5707 // Expand Worklist in topological order: whenever a new instruction
5708 // is added , its users should be either already inside Worklist, or
5709 // out of scope. It ensures a uniform instruction will only be used
5710 // by uniform instructions or out of scope instructions.
5711 unsigned idx = 0;
5712 while (idx != Worklist.size()) {
5713 Instruction *I = Worklist[idx++];
5714
5715 for (auto OV : I->operand_values()) {
5716 if (isOutOfScope(OV))
5717 continue;
5718 auto *OI = cast<Instruction>(OV);
5719 if (all_of(OI->users(), [&](User *U) -> bool {
5720 auto *J = cast<Instruction>(U);
5721 return !TheLoop->contains(J) || Worklist.count(J) ||
5722 (OI == getPointerOperand(J) && isUniformDecision(J, VF));
5723 })) {
5724 Worklist.insert(OI);
5725 DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: "
<< *OI << "\n"; } } while (false)
;
5726 }
5727 }
5728 }
5729
5730 // Returns true if Ptr is the pointer operand of a memory access instruction
5731 // I, and I is known to not require scalarization.
5732 auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
5733 return getPointerOperand(I) == Ptr && isUniformDecision(I, VF);
5734 };
5735
5736 // For an instruction to be added into Worklist above, all its users inside
5737 // the loop should also be in Worklist. However, this condition cannot be
5738 // true for phi nodes that form a cyclic dependence. We must process phi
5739 // nodes separately. An induction variable will remain uniform if all users
5740 // of the induction variable and induction variable update remain uniform.
5741 // The code below handles both pointer and non-pointer induction variables.
5742 for (auto &Induction : *Legal->getInductionVars()) {
5743 auto *Ind = Induction.first;
5744 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5745
5746 // Determine if all users of the induction variable are uniform after
5747 // vectorization.
5748 auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
5749 auto *I = cast<Instruction>(U);
5750 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
5751 isVectorizedMemAccessUse(I, Ind);
5752 });
5753 if (!UniformInd)
5754 continue;
5755
5756 // Determine if all users of the induction variable update instruction are
5757 // uniform after vectorization.
5758 auto UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5759 auto *I = cast<Instruction>(U);
5760 return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||
5761 isVectorizedMemAccessUse(I, IndUpdate);
5762 });
5763 if (!UniformIndUpdate)
5764 continue;
5765
5766 // The induction variable and its update instruction will remain uniform.
5767 Worklist.insert(Ind);
5768 Worklist.insert(IndUpdate);
5769 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Ind << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: "
<< *Ind << "\n"; } } while (false)
;
5770 DEBUG(dbgs() << "LV: Found uniform instruction: " << *IndUpdate << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: "
<< *IndUpdate << "\n"; } } while (false)
;
5771 }
5772
5773 Uniforms[VF].insert(Worklist.begin(), Worklist.end());
5774}
5775
5776bool LoopVectorizationLegality::canVectorizeMemory() {
5777 LAI = &(*GetLAA)(*TheLoop);
5778 InterleaveInfo.setLAI(LAI);
5779 const OptimizationRemarkAnalysis *LAR = LAI->getReport();
5780 if (LAR) {
5781 OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
5782 "loop not vectorized: ", *LAR);
5783 ORE->emit(VR);
5784 }
5785 if (!LAI->canVectorizeMemory())
5786 return false;
5787
5788 if (LAI->hasStoreToLoopInvariantAddress()) {
5789 ORE->emit(createMissedAnalysis("CantVectorizeStoreToLoopInvariantAddress")
5790 << "write to a loop invariant address could not be vectorized");
5791 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 (false)
;
5792 return false;
5793 }
5794
5795 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
5796 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
5797
5798 return true;
5799}
5800
5801bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5802 Value *In0 = const_cast<Value *>(V);
5803 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5804 if (!PN)
5805 return false;
5806
5807 return Inductions.count(PN);
5808}
5809
5810bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
5811 return FirstOrderRecurrences.count(Phi);
5812}
5813
5814bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5815 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
5816}
5817
5818bool LoopVectorizationLegality::blockCanBePredicated(
5819 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
5820 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
5821
5822 for (Instruction &I : *BB) {
5823 // Check that we don't have a constant expression that can trap as operand.
5824 for (Value *Operand : I.operands()) {
5825 if (auto *C = dyn_cast<Constant>(Operand))
5826 if (C->canTrap())
5827 return false;
5828 }
5829 // We might be able to hoist the load.
5830 if (I.mayReadFromMemory()) {
5831 auto *LI = dyn_cast<LoadInst>(&I);
5832 if (!LI)
5833 return false;
5834 if (!SafePtrs.count(LI->getPointerOperand())) {
5835 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
5836 isLegalMaskedGather(LI->getType())) {
5837 MaskedOp.insert(LI);
5838 continue;
5839 }
5840 // !llvm.mem.parallel_loop_access implies if-conversion safety.
5841 if (IsAnnotatedParallel)
5842 continue;
5843 return false;
5844 }
5845 }
5846
5847 if (I.mayWriteToMemory()) {
5848 auto *SI = dyn_cast<StoreInst>(&I);
5849 // We only support predication of stores in basic blocks with one
5850 // predecessor.
5851 if (!SI)
5852 return false;
5853
5854 // Build a masked store if it is legal for the target.
5855 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
5856 SI->getPointerOperand()) ||
5857 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
5858 MaskedOp.insert(SI);
5859 continue;
5860 }
5861
5862 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5863 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5864
5865 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5866 !isSinglePredecessor)
5867 return false;
5868 }
5869 if (I.mayThrow())
5870 return false;
5871 }
5872
5873 return true;
5874}
5875
5876void InterleavedAccessInfo::collectConstStrideAccesses(
5877 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
5878 const ValueToValueMap &Strides) {
5879
5880 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
5881
5882 // Since it's desired that the load/store instructions be maintained in
5883 // "program order" for the interleaved access analysis, we have to visit the
5884 // blocks in the loop in reverse postorder (i.e., in a topological order).
5885 // Such an ordering will ensure that any load/store that may be executed
5886 // before a second load/store will precede the second load/store in
5887 // AccessStrideInfo.
5888 LoopBlocksDFS DFS(TheLoop);
5889 DFS.perform(LI);
5890 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
5891 for (auto &I : *BB) {
5892 auto *LI = dyn_cast<LoadInst>(&I);
5893 auto *SI = dyn_cast<StoreInst>(&I);
5894 if (!LI && !SI)
5895 continue;
5896
5897 Value *Ptr = getPointerOperand(&I);
5898 // We don't check wrapping here because we don't know yet if Ptr will be
5899 // part of a full group or a group with gaps. Checking wrapping for all
5900 // pointers (even those that end up in groups with no gaps) will be overly
5901 // conservative. For full groups, wrapping should be ok since if we would
5902 // wrap around the address space we would do a memory access at nullptr
5903 // even without the transformation. The wrapping checks are therefore
5904 // deferred until after we've formed the interleaved groups.
5905 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
5906 /*Assume=*/true, /*ShouldCheckWrap=*/false);
5907
5908 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
5909 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5910 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
5911
5912 // An alignment of 0 means target ABI alignment.
5913 unsigned Align = getMemInstAlignment(&I);
5914 if (!Align)
5915 Align = DL.getABITypeAlignment(PtrTy->getElementType());
5916
5917 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
5918 }
5919}
5920
5921// Analyze interleaved accesses and collect them into interleaved load and
5922// store groups.
5923//
5924// When generating code for an interleaved load group, we effectively hoist all
5925// loads in the group to the location of the first load in program order. When
5926// generating code for an interleaved store group, we sink all stores to the
5927// location of the last store. This code motion can change the order of load
5928// and store instructions and may break dependences.
5929//
5930// The code generation strategy mentioned above ensures that we won't violate
5931// any write-after-read (WAR) dependences.
5932//
5933// E.g., for the WAR dependence: a = A[i]; // (1)
5934// A[i] = b; // (2)
5935//
5936// The store group of (2) is always inserted at or below (2), and the load
5937// group of (1) is always inserted at or above (1). Thus, the instructions will
5938// never be reordered. All other dependences are checked to ensure the
5939// correctness of the instruction reordering.
5940//
5941// The algorithm visits all memory accesses in the loop in bottom-up program
5942// order. Program order is established by traversing the blocks in the loop in
5943// reverse postorder when collecting the accesses.
5944//
5945// We visit the memory accesses in bottom-up order because it can simplify the
5946// construction of store groups in the presence of write-after-write (WAW)
5947// dependences.
5948//
5949// E.g., for the WAW dependence: A[i] = a; // (1)
5950// A[i] = b; // (2)
5951// A[i + 1] = c; // (3)
5952//
5953// We will first create a store group with (3) and (2). (1) can't be added to
5954// this group because it and (2) are dependent. However, (1) can be grouped
5955// with other accesses that may precede it in program order. Note that a
5956// bottom-up order does not imply that WAW dependences should not be checked.
5957void InterleavedAccessInfo::analyzeInterleaving(
5958 const ValueToValueMap &Strides) {
5959 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Analyzing interleaved accesses...\n"
; } } while (false)
;
5960
5961 // Holds all accesses with a constant stride.
5962 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
5963 collectConstStrideAccesses(AccessStrideInfo, Strides);
5964
5965 if (AccessStrideInfo.empty())
1
Assuming the condition is false
2
Taking false branch
5966 return;
5967
5968 // Collect the dependences in the loop.
5969 collectDependences();
5970
5971 // Holds all interleaved store groups temporarily.
5972 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
5973 // Holds all interleaved load groups temporarily.
5974 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
5975
5976 // Search in bottom-up program order for pairs of accesses (A and B) that can
5977 // form interleaved load or store groups. In the algorithm below, access A
5978 // precedes access B in program order. We initialize a group for B in the
5979 // outer loop of the algorithm, and then in the inner loop, we attempt to
5980 // insert each A into B's group if:
5981 //
5982 // 1. A and B have the same stride,
5983 // 2. A and B have the same memory object size, and
5984 // 3. A belongs in B's group according to its distance from B.
5985 //
5986 // Special care is taken to ensure group formation will not break any
5987 // dependences.
5988 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
3
Loop condition is true. Entering loop body
5989 BI != E; ++BI) {
5990 Instruction *B = BI->first;
5991 StrideDescriptor DesB = BI->second;
5992
5993 // Initialize a group for B if it has an allowable stride. Even if we don't
5994 // create a group for B, we continue with the bottom-up algorithm to ensure
5995 // we don't break any of B's dependences.
5996 InterleaveGroup *Group = nullptr;
4
'Group' initialized to a null pointer value
5997 if (isStrided(DesB.Stride)) {
5
Taking false branch
5998 Group = getInterleaveGroup(B);
5999 if (!Group) {
6000 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Creating an interleave group with:"
<< *B << '\n'; } } while (false)
;
6001 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
6002 }
6003 if (B->mayWriteToMemory())
6004 StoreGroups.insert(Group);
6005 else
6006 LoadGroups.insert(Group);
6007 }
6008
6009 for (auto AI = std::next(BI); AI != E; ++AI) {
6
Loop condition is true. Entering loop body
9
Loop condition is true. Entering loop body
12
Loop condition is true. Entering loop body
15
Loop condition is true. Entering loop body
6010 Instruction *A = AI->first;
6011 StrideDescriptor DesA = AI->second;
6012
6013 // Our code motion strategy implies that we can't have dependences
6014 // between accesses in an interleaved group and other accesses located
6015 // between the first and last member of the group. Note that this also
6016 // means that a group can't have more than one member at a given offset.
6017 // The accesses in a group can have dependences with other accesses, but
6018 // we must ensure we don't extend the boundaries of the group such that
6019 // we encompass those dependent accesses.
6020 //
6021 // For example, assume we have the sequence of accesses shown below in a
6022 // stride-2 loop:
6023 //
6024 // (1, 2) is a group | A[i] = a; // (1)
6025 // | A[i-1] = b; // (2) |
6026 // A[i-3] = c; // (3)
6027 // A[i] = d; // (4) | (2, 4) is not a group
6028 //
6029 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
6030 // but not with (4). If we did, the dependent access (3) would be within
6031 // the boundaries of the (2, 4) group.
6032 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
7
Taking false branch
10
Taking false branch
13
Taking false branch
16
Taking false branch
6033
6034 // If a dependence exists and A is already in a group, we know that A
6035 // must be a store since A precedes B and WAR dependences are allowed.
6036 // Thus, A would be sunk below B. We release A's group to prevent this
6037 // illegal code motion. A will then be free to form another group with
6038 // instructions that precede it.
6039 if (isInterleaved(A)) {
6040 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
6041 StoreGroups.remove(StoreGroup);
6042 releaseGroup(StoreGroup);
6043 }
6044
6045 // If a dependence exists and A is not already in a group (or it was
6046 // and we just released it), B might be hoisted above A (if B is a
6047 // load) or another store might be sunk below A (if B is a store). In
6048 // either case, we can't add additional instructions to B's group. B
6049 // will only form a group with instructions that it precedes.
6050 break;
6051 }
6052
6053 // At this point, we've checked for illegal code motion. If either A or B
6054 // isn't strided, there's nothing left to do.
6055 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
17
Taking false branch
6056 continue;
8
Execution continues on line 6009
11
Execution continues on line 6009
14
Execution continues on line 6009
6057
6058 // Ignore A if it's already in a group or isn't the same kind of memory
6059 // operation as B.
6060 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
18
Assuming the condition is false
19
Taking false branch
6061 continue;
6062
6063 // Check rules 1 and 2. Ignore A if its stride or size is different from
6064 // that of B.
6065 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
20
Taking false branch
6066 continue;
6067
6068 // Ignore A if the memory object of A and B don't belong to the same
6069 // address space
6070 if (getMemInstAddressSpace(A) != getMemInstAddressSpace(B))
21
Taking false branch
6071 continue;
6072
6073 // Calculate the distance from A to B.
6074 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
6075 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
6076 if (!DistToB)
22
Assuming 'DistToB' is non-null
23
Taking false branch
6077 continue;
6078 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
6079
6080 // Check rule 3. Ignore A if its distance to B is not a multiple of the
6081 // size.
6082 if (DistanceToB % static_cast<int64_t>(DesB.Size))
24
Taking false branch
6083 continue;
6084
6085 // Ignore A if either A or B is in a predicated block. Although we
6086 // currently prevent group formation for predicated accesses, we may be
6087 // able to relax this limitation in the future once we handle more
6088 // complicated blocks.
6089 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
25
Assuming the condition is false
26
Assuming the condition is false
27
Taking false branch
6090 continue;
6091
6092 // The index of A is the index of B plus A's distance to B in multiples
6093 // of the size.
6094 int IndexA =
6095 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
28
Called C++ object pointer is null
6096
6097 // Try to insert A into B's group.
6098 if (Group->insertMember(A, IndexA, DesA.Align)) {
6099 DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Inserted:" <<
*A << '\n' << " into the interleave group with"
<< *B << '\n'; } } while (false)
6100 << " into the interleave group with" << *B << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Inserted:" <<
*A << '\n' << " into the interleave group with"
<< *B << '\n'; } } while (false)
;
6101 InterleaveGroupMap[A] = Group;
6102
6103 // Set the first load in program order as the insert position.
6104 if (A->mayReadFromMemory())
6105 Group->setInsertPos(A);
6106 }
6107 } // Iteration over A accesses.
6108 } // Iteration over B accesses.
6109
6110 // Remove interleaved store groups with gaps.
6111 for (InterleaveGroup *Group : StoreGroups)
6112 if (Group->getNumMembers() != Group->getFactor())
6113 releaseGroup(Group);
6114
6115 // Remove interleaved groups with gaps (currently only loads) whose memory
6116 // accesses may wrap around. We have to revisit the getPtrStride analysis,
6117 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
6118 // not check wrapping (see documentation there).
6119 // FORNOW we use Assume=false;
6120 // TODO: Change to Assume=true but making sure we don't exceed the threshold
6121 // of runtime SCEV assumptions checks (thereby potentially failing to
6122 // vectorize altogether).
6123 // Additional optional optimizations:
6124 // TODO: If we are peeling the loop and we know that the first pointer doesn't
6125 // wrap then we can deduce that all pointers in the group don't wrap.
6126 // This means that we can forcefully peel the loop in order to only have to
6127 // check the first pointer for no-wrap. When we'll change to use Assume=true
6128 // we'll only need at most one runtime check per interleaved group.
6129 //
6130 for (InterleaveGroup *Group : LoadGroups) {
6131
6132 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
6133 // load would wrap around the address space we would do a memory access at
6134 // nullptr even without the transformation.
6135 if (Group->getNumMembers() == Group->getFactor())
6136 continue;
6137
6138 // Case 2: If first and last members of the group don't wrap this implies
6139 // that all the pointers in the group don't wrap.
6140 // So we check only group member 0 (which is always guaranteed to exist),
6141 // and group member Factor - 1; If the latter doesn't exist we rely on
6142 // peeling (if it is a non-reveresed accsess -- see Case 3).
6143 Value *FirstMemberPtr = getPointerOperand(Group->getMember(0));
6144 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
6145 /*ShouldCheckWrap=*/true)) {
6146 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"first group member potentially pointer-wrapping.\n"; } } while
(false)
6147 "first group member potentially pointer-wrapping.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"first group member potentially pointer-wrapping.\n"; } } while
(false)
;
6148 releaseGroup(Group);
6149 continue;
6150 }
6151 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
6152 if (LastMember) {
6153 Value *LastMemberPtr = getPointerOperand(LastMember);
6154 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
6155 /*ShouldCheckWrap=*/true)) {
6156 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"last group member potentially pointer-wrapping.\n"; } } while
(false)
6157 "last group member potentially pointer-wrapping.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"last group member potentially pointer-wrapping.\n"; } } while
(false)
;
6158 releaseGroup(Group);
6159 }
6160 } else {
6161 // Case 3: A non-reversed interleaved load group with gaps: We need
6162 // to execute at least one scalar epilogue iteration. This will ensure
6163 // we don't speculatively access memory out-of-bounds. We only need
6164 // to look for a member at index factor - 1, since every group must have
6165 // a member at index zero.
6166 if (Group->isReverse()) {
6167 releaseGroup(Group);
6168 continue;
6169 }
6170 DEBUG(dbgs() << "LV: Interleaved group requires epilogue iteration.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Interleaved group requires epilogue iteration.\n"
; } } while (false)
;
6171 RequiresScalarEpilogue = true;
6172 }
6173 }
6174}
6175
6176Optional<unsigned> LoopVectorizationCostModel::computeMaxVF(bool OptForSize) {
6177 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
6178 ORE->emit(createMissedAnalysis("ConditionalStore")
6179 << "store that is conditionally executed prevents vectorization");
6180 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 (false)
;
6181 return None;
6182 }
6183
6184 if (!OptForSize) // Remaining checks deal with scalar loop when OptForSize.
6185 return computeFeasibleMaxVF(OptForSize);
6186
6187 if (Legal->getRuntimePointerChecking()->Need) {
6188 ORE->emit(createMissedAnalysis("CantVersionLoopWithOptForSize")
6189 << "runtime pointer checks needed. Enable vectorization of this "
6190 "loop with '#pragma clang loop vectorize(enable)' when "
6191 "compiling with -Os/-Oz");
6192 DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n"
; } } while (false)
6193 << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n"
; } } while (false)
;
6194 return None;
6195 }
6196
6197 // If we optimize the program for size, avoid creating the tail loop.
6198 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6199 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found trip count: "
<< TC << '\n'; } } while (false)
;
6200
6201 // If we don't know the precise trip count, don't try to vectorize.
6202 if (TC < 2) {
6203 ORE->emit(
6204 createMissedAnalysis("UnknownLoopCountComplexCFG")
6205 << "unable to calculate the loop count due to complex control flow");
6206 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n"
; } } while (false)
;
6207 return None;
6208 }
6209
6210 unsigned MaxVF = computeFeasibleMaxVF(OptForSize);
6211
6212 if (TC % MaxVF != 0) {
6213 // If the trip count that we found modulo the vectorization factor is not
6214 // zero then we require a tail.
6215 // FIXME: look for a smaller MaxVF that does divide TC rather than give up.
6216 // FIXME: return None if loop requiresScalarEpilog(<MaxVF>), or look for a
6217 // smaller MaxVF that does not require a scalar epilog.
6218
6219 ORE->emit(createMissedAnalysis("NoTailLoopWithOptForSize")
6220 << "cannot optimize for size and vectorize at the "
6221 "same time. Enable vectorization of this loop "
6222 "with '#pragma clang loop vectorize(enable)' "
6223 "when compiling with -Os/-Oz");
6224 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n"
; } } while (false)
;
6225 return None;
6226 }
6227
6228 return MaxVF;
6229}
6230
6231unsigned LoopVectorizationCostModel::computeFeasibleMaxVF(bool OptForSize) {
6232 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
6233 unsigned SmallestType, WidestType;
6234 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
6235 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
6236 unsigned MaxSafeDepDist = -1U;
6237
6238 // Get the maximum safe dependence distance in bits computed by LAA. If the
6239 // loop contains any interleaved accesses, we divide the dependence distance
6240 // by the maximum interleave factor of all interleaved groups. Note that
6241 // although the division ensures correctness, this is a fairly conservative
6242 // computation because the maximum distance computed by LAA may not involve
6243 // any of the interleaved accesses.
6244 if (Legal->getMaxSafeDepDistBytes() != -1U)
6245 MaxSafeDepDist =
6246 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
6247
6248 WidestRegister =
6249 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
6250 unsigned MaxVectorSize = WidestRegister / WidestType;
6251
6252 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: "
<< SmallestType << " / " << WidestType <<
" bits.\n"; } } while (false)
6253 << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: "
<< SmallestType << " / " << WidestType <<
" bits.\n"; } } while (false)
;
6254 DEBUG(dbgs() << "LV: The Widest register is: " << WidestRegisterdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest register is: "
<< WidestRegister << " bits.\n"; } } while (false
)
6255 << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest register is: "
<< WidestRegister << " bits.\n"; } } while (false
)
;
6256
6257 if (MaxVectorSize == 0) {
6258 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 (false)
;
6259 MaxVectorSize = 1;
6260 }
6261
6262 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6263, __PRETTY_FUNCTION__))
6263 " 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6263, __PRETTY_FUNCTION__))
;
6264
6265 unsigned MaxVF = MaxVectorSize;
6266 if (MaximizeBandwidth && !OptForSize) {
6267 // Collect all viable vectorization factors.
6268 SmallVector<unsigned, 8> VFs;
6269 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
6270 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
6271 VFs.push_back(VS);
6272
6273 // For each VF calculate its register usage.
6274 auto RUs = calculateRegisterUsage(VFs);
6275
6276 // Select the largest VF which doesn't require more registers than existing
6277 // ones.
6278 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
6279 for (int i = RUs.size() - 1; i >= 0; --i) {
6280 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
6281 MaxVF = VFs[i];
6282 break;
6283 }
6284 }
6285 }
6286 return MaxVF;
6287}
6288
6289LoopVectorizationCostModel::VectorizationFactor
6290LoopVectorizationCostModel::selectVectorizationFactor(unsigned MaxVF) {
6291 float Cost = expectedCost(1).first;
6292#ifndef NDEBUG
6293 const float ScalarCost = Cost;
6294#endif /* NDEBUG */
6295 unsigned Width = 1;
6296 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 (false)
;
6297
6298 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
6299 // Ignore scalar width, because the user explicitly wants vectorization.
6300 if (ForceVectorization && MaxVF > 1) {
6301 Width = 2;
6302 Cost = expectedCost(Width).first / (float)Width;
6303 }
6304
6305 for (unsigned i = 2; i <= MaxVF; i *= 2) {
6306 // Notice that the vector loop needs to be executed less times, so
6307 // we need to divide the cost of the vector loops by the width of
6308 // the vector elements.
6309 VectorizationCostTy C = expectedCost(i);
6310 float VectorCost = C.first / (float)i;
6311 DEBUG(dbgs() << "LV: Vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Vector loop of width "
<< i << " costs: " << (int)VectorCost <<
".\n"; } } while (false)
6312 << " costs: " << (int)VectorCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Vector loop of width "
<< i << " costs: " << (int)VectorCost <<
".\n"; } } while (false)
;
6313 if (!C.second && !ForceVectorization) {
6314 DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width "
<< i << " because it will not generate any vector instructions.\n"
; } } while (false)
6315 dbgs() << "LV: Not considering vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width "
<< i << " because it will not generate any vector instructions.\n"
; } } while (false)
6316 << " because it will not generate any vector instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width "
<< i << " because it will not generate any vector instructions.\n"
; } } while (false)
;
6317 continue;
6318 }
6319 if (VectorCost < Cost) {
6320 Cost = VectorCost;
6321 Width = i;
6322 }
6323 }
6324
6325 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 (false)
6326 << "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 (false)
6327 << "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 (false)
;
6328 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Selecting VF: " <<
Width << ".\n"; } } while (false)
;
6329 VectorizationFactor Factor = {Width, (unsigned)(Width * Cost)};
6330 return Factor;
6331}
6332
6333std::pair<unsigned, unsigned>
6334LoopVectorizationCostModel::getSmallestAndWidestTypes() {
6335 unsigned MinWidth = -1U;
6336 unsigned MaxWidth = 8;
6337 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6338
6339 // For each block.
6340 for (BasicBlock *BB : TheLoop->blocks()) {
6341 // For each instruction in the loop.
6342 for (Instruction &I : *BB) {
6343 Type *T = I.getType();
6344
6345 // Skip ignored values.
6346 if (ValuesToIgnore.count(&I))
6347 continue;
6348
6349 // Only examine Loads, Stores and PHINodes.
6350 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
6351 continue;
6352
6353 // Examine PHI nodes that are reduction variables. Update the type to
6354 // account for the recurrence type.
6355 if (auto *PN = dyn_cast<PHINode>(&I)) {
6356 if (!Legal->isReductionVariable(PN))
6357 continue;
6358 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
6359 T = RdxDesc.getRecurrenceType();
6360 }
6361
6362 // Examine the stored values.
6363 if (auto *ST = dyn_cast<StoreInst>(&I))
6364 T = ST->getValueOperand()->getType();
6365
6366 // Ignore loaded pointer types and stored pointer types that are not
6367 // vectorizable.
6368 //
6369 // FIXME: The check here attempts to predict whether a load or store will
6370 // be vectorized. We only know this for certain after a VF has
6371 // been selected. Here, we assume that if an access can be
6372 // vectorized, it will be. We should also look at extending this
6373 // optimization to non-pointer types.
6374 //
6375 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I) &&
6376 !Legal->isAccessInterleaved(&I) && !Legal->isLegalGatherOrScatter(&I))
6377 continue;
6378
6379 MinWidth = std::min(MinWidth,
6380 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6381 MaxWidth = std::max(MaxWidth,
6382 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6383 }
6384 }
6385
6386 return {MinWidth, MaxWidth};
6387}
6388
6389unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
6390 unsigned VF,
6391 unsigned LoopCost) {
6392
6393 // -- The interleave heuristics --
6394 // We interleave the loop in order to expose ILP and reduce the loop overhead.
6395 // There are many micro-architectural considerations that we can't predict
6396 // at this level. For example, frontend pressure (on decode or fetch) due to
6397 // code size, or the number and capabilities of the execution ports.
6398 //
6399 // We use the following heuristics to select the interleave count:
6400 // 1. If the code has reductions, then we interleave to break the cross
6401 // iteration dependency.
6402 // 2. If the loop is really small, then we interleave to reduce the loop
6403 // overhead.
6404 // 3. We don't interleave if we think that we will spill registers to memory
6405 // due to the increased register pressure.
6406
6407 // When we optimize for size, we don't interleave.
6408 if (OptForSize)
6409 return 1;
6410
6411 // We used the distance for the interleave count.
6412 if (Legal->getMaxSafeDepDistBytes() != -1U)
6413 return 1;
6414
6415 // Do not interleave loops with a relatively small trip count.
6416 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6417 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
6418 return 1;
6419
6420 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
6421 DEBUG(dbgs() << "LV: The target has " << TargetNumRegistersdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (false
)
6422 << " registers\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (false
)
;
6423
6424 if (VF == 1) {
6425 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
6426 TargetNumRegisters = ForceTargetNumScalarRegs;
6427 } else {
6428 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
6429 TargetNumRegisters = ForceTargetNumVectorRegs;
6430 }
6431
6432 RegisterUsage R = calculateRegisterUsage({VF})[0];
6433 // We divide by these constants so assume that we have at least one
6434 // instruction that uses at least one register.
6435 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
6436 R.NumInstructions = std::max(R.NumInstructions, 1U);
6437
6438 // We calculate the interleave count using the following formula.
6439 // Subtract the number of loop invariants from the number of available
6440 // registers. These registers are used by all of the interleaved instances.
6441 // Next, divide the remaining registers by the number of registers that is
6442 // required by the loop, in order to estimate how many parallel instances
6443 // fit without causing spills. All of this is rounded down if necessary to be
6444 // a power of two. We want power of two interleave count to simplify any
6445 // addressing operations or alignment considerations.
6446 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
6447 R.MaxLocalUsers);
6448
6449 // Don't count the induction variable as interleaved.
6450 if (EnableIndVarRegisterHeur)
6451 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
6452 std::max(1U, (R.MaxLocalUsers - 1)));
6453
6454 // Clamp the interleave ranges to reasonable counts.
6455 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
6456
6457 // Check if the user has overridden the max.
6458 if (VF == 1) {
6459 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
6460 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
6461 } else {
6462 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
6463 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
6464 }
6465
6466 // If we did not calculate the cost for VF (because the user selected the VF)
6467 // then we calculate the cost of VF here.
6468 if (LoopCost == 0)
6469 LoopCost = expectedCost(VF).first;
6470
6471 // Clamp the calculated IC to be between the 1 and the max interleave count
6472 // that the target allows.
6473 if (IC > MaxInterleaveCount)
6474 IC = MaxInterleaveCount;
6475 else if (IC < 1)
6476 IC = 1;
6477
6478 // Interleave if we vectorized this loop and there is a reduction that could
6479 // benefit from interleaving.
6480 if (VF > 1 && Legal->getReductionVars()->size()) {
6481 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Interleaving because of reductions.\n"
; } } while (false)
;
6482 return IC;
6483 }
6484
6485 // Note that if we've already vectorized the loop we will have done the
6486 // runtime check and so interleaving won't require further checks.
6487 bool InterleavingRequiresRuntimePointerCheck =
6488 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
6489
6490 // We want to interleave small loops in order to reduce the loop overhead and
6491 // potentially expose ILP opportunities.
6492 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Loop cost is " <<
LoopCost << '\n'; } } while (false)
;
6493 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
6494 // We assume that the cost overhead is 1 and we use the cost model
6495 // to estimate the cost of the loop and interleave until the cost of the
6496 // loop overhead is about 5% of the cost of the loop.
6497 unsigned SmallIC =
6498 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
6499
6500 // Interleave until store/load ports (estimated by max interleave count) are
6501 // saturated.
6502 unsigned NumStores = Legal->getNumStores();
6503 unsigned NumLoads = Legal->getNumLoads();
6504 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
6505 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
6506
6507 // If we have a scalar reduction (vector reductions are already dealt with
6508 // by this point), we can increase the critical path length if the loop
6509 // we're interleaving is inside another loop. Limit, by default to 2, so the
6510 // critical path only gets increased by one reduction operation.
6511 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
6512 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
6513 SmallIC = std::min(SmallIC, F);
6514 StoresIC = std::min(StoresIC, F);
6515 LoadsIC = std::min(LoadsIC, F);
6516 }
6517
6518 if (EnableLoadStoreRuntimeInterleave &&
6519 std::max(StoresIC, LoadsIC) > SmallIC) {
6520 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n"
; } } while (false)
;
6521 return std::max(StoresIC, LoadsIC);
6522 }
6523
6524 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Interleaving to reduce branch cost.\n"
; } } while (false)
;
6525 return SmallIC;
6526 }
6527
6528 // Interleave if this is a large loop (small loops are already dealt with by
6529 // this point) that could benefit from interleaving.
6530 bool HasReductions = (Legal->getReductionVars()->size() > 0);
6531 if (TTI.enableAggressiveInterleaving(HasReductions)) {
6532 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Interleaving to expose ILP.\n"
; } } while (false)
;
6533 return IC;
6534 }
6535
6536 DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n"
; } } while (false)
;
6537 return 1;
6538}
6539
6540SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
6541LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
6542 // This function calculates the register usage by measuring the highest number
6543 // of values that are alive at a single location. Obviously, this is a very
6544 // rough estimation. We scan the loop in a topological order in order and
6545 // assign a number to each instruction. We use RPO to ensure that defs are
6546 // met before their users. We assume that each instruction that has in-loop
6547 // users starts an interval. We record every time that an in-loop value is
6548 // used, so we have a list of the first and last occurrences of each
6549 // instruction. Next, we transpose this data structure into a multi map that
6550 // holds the list of intervals that *end* at a specific location. This multi
6551 // map allows us to perform a linear search. We scan the instructions linearly
6552 // and record each time that a new interval starts, by placing it in a set.
6553 // If we find this value in the multi-map then we remove it from the set.
6554 // The max register usage is the maximum size of the set.
6555 // We also search for instructions that are defined outside the loop, but are
6556 // used inside the loop. We need this number separately from the max-interval
6557 // usage number because when we unroll, loop-invariant values do not take
6558 // more register.
6559 LoopBlocksDFS DFS(TheLoop);
6560 DFS.perform(LI);
6561
6562 RegisterUsage RU;
6563 RU.NumInstructions = 0;
6564
6565 // Each 'key' in the map opens a new interval. The values
6566 // of the map are the index of the 'last seen' usage of the
6567 // instruction that is the key.
6568 typedef DenseMap<Instruction *, unsigned> IntervalMap;
6569 // Maps instruction to its index.
6570 DenseMap<unsigned, Instruction *> IdxToInstr;
6571 // Marks the end of each interval.
6572 IntervalMap EndPoint;
6573 // Saves the list of instruction indices that are used in the loop.
6574 SmallSet<Instruction *, 8> Ends;
6575 // Saves the list of values that are used in the loop but are
6576 // defined outside the loop, such as arguments and constants.
6577 SmallPtrSet<Value *, 8> LoopInvariants;
6578
6579 unsigned Index = 0;
6580 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
6581 RU.NumInstructions += BB->size();
6582 for (Instruction &I : *BB) {
6583 IdxToInstr[Index++] = &I;
6584
6585 // Save the end location of each USE.
6586 for (Value *U : I.operands()) {
6587 auto *Instr = dyn_cast<Instruction>(U);
6588
6589 // Ignore non-instruction values such as arguments, constants, etc.
6590 if (!Instr)
6591 continue;
6592
6593 // If this instruction is outside the loop then record it and continue.
6594 if (!TheLoop->contains(Instr)) {
6595 LoopInvariants.insert(Instr);
6596 continue;
6597 }
6598
6599 // Overwrite previous end points.
6600 EndPoint[Instr] = Index;
6601 Ends.insert(Instr);
6602 }
6603 }
6604 }
6605
6606 // Saves the list of intervals that end with the index in 'key'.
6607 typedef SmallVector<Instruction *, 2> InstrList;
6608 DenseMap<unsigned, InstrList> TransposeEnds;
6609
6610 // Transpose the EndPoints to a list of values that end at each index.
6611 for (auto &Interval : EndPoint)
6612 TransposeEnds[Interval.second].push_back(Interval.first);
6613
6614 SmallSet<Instruction *, 8> OpenIntervals;
6615
6616 // Get the size of the widest register.
6617 unsigned MaxSafeDepDist = -1U;
6618 if (Legal->getMaxSafeDepDistBytes() != -1U)
6619 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
6620 unsigned WidestRegister =
6621 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
6622 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6623
6624 SmallVector<RegisterUsage, 8> RUs(VFs.size());
6625 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
6626
6627 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 (false)
;
6628
6629 // A lambda that gets the register usage for the given type and VF.
6630 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
6631 if (Ty->isTokenTy())
6632 return 0U;
6633 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
6634 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
6635 };
6636
6637 for (unsigned int i = 0; i < Index; ++i) {
6638 Instruction *I = IdxToInstr[i];
6639
6640 // Remove all of the instructions that end at this location.
6641 InstrList &List = TransposeEnds[i];
6642 for (Instruction *ToRemove : List)
6643 OpenIntervals.erase(ToRemove);
6644
6645 // Ignore instructions that are never used within the loop.
6646 if (!Ends.count(I))
6647 continue;
6648
6649 // Skip ignored values.
6650 if (ValuesToIgnore.count(I))
6651 continue;
6652
6653 // For each VF find the maximum usage of registers.
6654 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
6655 if (VFs[j] == 1) {
6656 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
6657 continue;
6658 }
6659 collectUniformsAndScalars(VFs[j]);
6660 // Count the number of live intervals.
6661 unsigned RegUsage = 0;
6662 for (auto Inst : OpenIntervals) {
6663 // Skip ignored values for VF > 1.
6664 if (VecValuesToIgnore.count(Inst) ||
6665 isScalarAfterVectorization(Inst, VFs[j]))
6666 continue;
6667 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
6668 }
6669 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
6670 }
6671
6672 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 (false)
6673 << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): At #" <<
i << " Interval # " << OpenIntervals.size() <<
'\n'; } } while (false)
;
6674
6675 // Add the current instruction to the list of open intervals.
6676 OpenIntervals.insert(I);
6677 }
6678
6679 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
6680 unsigned Invariant = 0;
6681 if (VFs[i] == 1)
6682 Invariant = LoopInvariants.size();
6683 else {
6684 for (auto Inst : LoopInvariants)
6685 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
6686 }
6687
6688 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): VF = " <<
VFs[i] << '\n'; } } while (false)
;
6689 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): Found max usage: "
<< MaxUsages[i] << '\n'; } } while (false)
;
6690 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 (false)
;
6691 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): LoopSize: " <<
RU.NumInstructions << '\n'; } } while (false)
;
6692
6693 RU.LoopInvariantRegs = Invariant;
6694 RU.MaxLocalUsers = MaxUsages[i];
6695 RUs[i] = RU;
6696 }
6697
6698 return RUs;
6699}
6700
6701void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
6702
6703 // If we aren't vectorizing the loop, or if we've already collected the
6704 // instructions to scalarize, there's nothing to do. Collection may already
6705 // have occurred if we have a user-selected VF and are now computing the
6706 // expected cost for interleaving.
6707 if (VF < 2 || InstsToScalarize.count(VF))
6708 return;
6709
6710 // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
6711 // not profitable to scalarize any instructions, the presence of VF in the
6712 // map will indicate that we've analyzed it already.
6713 ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
6714
6715 // Find all the instructions that are scalar with predication in the loop and
6716 // determine if it would be better to not if-convert the blocks they are in.
6717 // If so, we also record the instructions to scalarize.
6718 for (BasicBlock *BB : TheLoop->blocks()) {
6719 if (!Legal->blockNeedsPredication(BB))
6720 continue;
6721 for (Instruction &I : *BB)
6722 if (Legal->isScalarWithPredication(&I)) {
6723 ScalarCostsTy ScalarCosts;
6724 if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
6725 ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
6726 }
6727 }
6728}
6729
6730int LoopVectorizationCostModel::computePredInstDiscount(
6731 Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
6732 unsigned VF) {
6733
6734 assert(!isUniformAfterVectorization(PredInst, VF) &&((!isUniformAfterVectorization(PredInst, VF) && "Instruction marked uniform-after-vectorization will be predicated"
) ? static_cast<void> (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6735, __PRETTY_FUNCTION__))
6735 "Instruction marked uniform-after-vectorization will be predicated")((!isUniformAfterVectorization(PredInst, VF) && "Instruction marked uniform-after-vectorization will be predicated"
) ? static_cast<void> (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6735, __PRETTY_FUNCTION__))
;
6736
6737 // Initialize the discount to zero, meaning that the scalar version and the
6738 // vector version cost the same.
6739 int Discount = 0;
6740
6741 // Holds instructions to analyze. The instructions we visit are mapped in
6742 // ScalarCosts. Those instructions are the ones that would be scalarized if
6743 // we find that the scalar version costs less.
6744 SmallVector<Instruction *, 8> Worklist;
6745
6746 // Returns true if the given instruction can be scalarized.
6747 auto canBeScalarized = [&](Instruction *I) -> bool {
6748
6749 // We only attempt to scalarize instructions forming a single-use chain
6750 // from the original predicated block that would otherwise be vectorized.
6751 // Although not strictly necessary, we give up on instructions we know will
6752 // already be scalar to avoid traversing chains that are unlikely to be
6753 // beneficial.
6754 if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
6755 isScalarAfterVectorization(I, VF))
6756 return false;
6757
6758 // If the instruction is scalar with predication, it will be analyzed
6759 // separately. We ignore it within the context of PredInst.
6760 if (Legal->isScalarWithPredication(I))
6761 return false;
6762
6763 // If any of the instruction's operands are uniform after vectorization,
6764 // the instruction cannot be scalarized. This prevents, for example, a
6765 // masked load from being scalarized.
6766 //
6767 // We assume we will only emit a value for lane zero of an instruction
6768 // marked uniform after vectorization, rather than VF identical values.
6769 // Thus, if we scalarize an instruction that uses a uniform, we would
6770 // create uses of values corresponding to the lanes we aren't emitting code
6771 // for. This behavior can be changed by allowing getScalarValue to clone
6772 // the lane zero values for uniforms rather than asserting.
6773 for (Use &U : I->operands())
6774 if (auto *J = dyn_cast<Instruction>(U.get()))
6775 if (isUniformAfterVectorization(J, VF))
6776 return false;
6777
6778 // Otherwise, we can scalarize the instruction.
6779 return true;
6780 };
6781
6782 // Returns true if an operand that cannot be scalarized must be extracted
6783 // from a vector. We will account for this scalarization overhead below. Note
6784 // that the non-void predicated instructions are placed in their own blocks,
6785 // and their return values are inserted into vectors. Thus, an extract would
6786 // still be required.
6787 auto needsExtract = [&](Instruction *I) -> bool {
6788 return TheLoop->contains(I) && !isScalarAfterVectorization(I, VF);
6789 };
6790
6791 // Compute the expected cost discount from scalarizing the entire expression
6792 // feeding the predicated instruction. We currently only consider expressions
6793 // that are single-use instruction chains.
6794 Worklist.push_back(PredInst);
6795 while (!Worklist.empty()) {
6796 Instruction *I = Worklist.pop_back_val();
6797
6798 // If we've already analyzed the instruction, there's nothing to do.
6799 if (ScalarCosts.count(I))
6800 continue;
6801
6802 // Compute the cost of the vector instruction. Note that this cost already
6803 // includes the scalarization overhead of the predicated instruction.
6804 unsigned VectorCost = getInstructionCost(I, VF).first;
6805
6806 // Compute the cost of the scalarized instruction. This cost is the cost of
6807 // the instruction as if it wasn't if-converted and instead remained in the
6808 // predicated block. We will scale this cost by block probability after
6809 // computing the scalarization overhead.
6810 unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
6811
6812 // Compute the scalarization overhead of needed insertelement instructions
6813 // and phi nodes.
6814 if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
6815 ScalarCost += TTI.getScalarizationOverhead(ToVectorTy(I->getType(), VF),
6816 true, false);
6817 ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
6818 }
6819
6820 // Compute the scalarization overhead of needed extractelement
6821 // instructions. For each of the instruction's operands, if the operand can
6822 // be scalarized, add it to the worklist; otherwise, account for the
6823 // overhead.
6824 for (Use &U : I->operands())
6825 if (auto *J = dyn_cast<Instruction>(U.get())) {
6826 assert(VectorType::isValidElementType(J->getType()) &&((VectorType::isValidElementType(J->getType()) && "Instruction has non-scalar type"
) ? static_cast<void> (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6827, __PRETTY_FUNCTION__))
6827 "Instruction has non-scalar type")((VectorType::isValidElementType(J->getType()) && "Instruction has non-scalar type"
) ? static_cast<void> (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6827, __PRETTY_FUNCTION__))
;
6828 if (canBeScalarized(J))
6829 Worklist.push_back(J);
6830 else if (needsExtract(J))
6831 ScalarCost += TTI.getScalarizationOverhead(
6832 ToVectorTy(J->getType(),VF), false, true);
6833 }
6834
6835 // Scale the total scalar cost by block probability.
6836 ScalarCost /= getReciprocalPredBlockProb();
6837
6838 // Compute the discount. A non-negative discount means the vector version
6839 // of the instruction costs more, and scalarizing would be beneficial.
6840 Discount += VectorCost - ScalarCost;
6841 ScalarCosts[I] = ScalarCost;
6842 }
6843
6844 return Discount;
6845}
6846
6847LoopVectorizationCostModel::VectorizationCostTy
6848LoopVectorizationCostModel::expectedCost(unsigned VF) {
6849 VectorizationCostTy Cost;
6850
6851 // Collect Uniform and Scalar instructions after vectorization with VF.
6852 collectUniformsAndScalars(VF);
6853
6854 // Collect the instructions (and their associated costs) that will be more
6855 // profitable to scalarize.
6856 collectInstsToScalarize(VF);
6857
6858 // For each block.
6859 for (BasicBlock *BB : TheLoop->blocks()) {
6860 VectorizationCostTy BlockCost;
6861
6862 // For each instruction in the old loop.
6863 for (Instruction &I : *BB) {
6864 // Skip dbg intrinsics.
6865 if (isa<DbgInfoIntrinsic>(I))
6866 continue;
6867
6868 // Skip ignored values.
6869 if (ValuesToIgnore.count(&I))
6870 continue;
6871
6872 VectorizationCostTy C = getInstructionCost(&I, VF);
6873
6874 // Check if we should override the cost.
6875 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
6876 C.first = ForceTargetInstructionCost;
6877
6878 BlockCost.first += C.first;
6879 BlockCost.second |= C.second;
6880 DEBUG(dbgs() << "LV: Found an estimated cost of " << C.first << " for VF "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of "
<< C.first << " for VF " << VF << " For instruction: "
<< I << '\n'; } } while (false)
6881 << VF << " For instruction: " << I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of "
<< C.first << " for VF " << VF << " For instruction: "
<< I << '\n'; } } while (false)
;
6882 }
6883
6884 // If we are vectorizing a predicated block, it will have been
6885 // if-converted. This means that the block's instructions (aside from
6886 // stores and instructions that may divide by zero) will now be
6887 // unconditionally executed. For the scalar case, we may not always execute
6888 // the predicated block. Thus, scale the block's cost by the probability of
6889 // executing it.
6890 if (VF == 1 && Legal->blockNeedsPredication(BB))
6891 BlockCost.first /= getReciprocalPredBlockProb();
6892
6893 Cost.first += BlockCost.first;
6894 Cost.second |= BlockCost.second;
6895 }
6896
6897 return Cost;
6898}
6899
6900/// \brief Gets Address Access SCEV after verifying that the access pattern
6901/// is loop invariant except the induction variable dependence.
6902///
6903/// This SCEV can be sent to the Target in order to estimate the address
6904/// calculation cost.
6905static const SCEV *getAddressAccessSCEV(
6906 Value *Ptr,
6907 LoopVectorizationLegality *Legal,
6908 ScalarEvolution *SE,
6909 const Loop *TheLoop) {
6910 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
6911 if (!Gep)
6912 return nullptr;
6913
6914 // We are looking for a gep with all loop invariant indices except for one
6915 // which should be an induction variable.
6916 unsigned NumOperands = Gep->getNumOperands();
6917 for (unsigned i = 1; i < NumOperands; ++i) {
6918 Value *Opd = Gep->getOperand(i);
6919 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
6920 !Legal->isInductionVariable(Opd))
6921 return nullptr;
6922 }
6923
6924 // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.
6925 return SE->getSCEV(Ptr);
6926}
6927
6928static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
6929 return Legal->hasStride(I->getOperand(0)) ||
6930 Legal->hasStride(I->getOperand(1));
6931}
6932
6933unsigned LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,
6934 unsigned VF) {
6935 Type *ValTy = getMemInstValueType(I);
6936 auto SE = PSE.getSE();
6937
6938 unsigned Alignment = getMemInstAlignment(I);
6939 unsigned AS = getMemInstAddressSpace(I);
6940 Value *Ptr = getPointerOperand(I);
6941 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6942
6943 // Figure out whether the access is strided and get the stride value
6944 // if it's known in compile time
6945 const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop);
6946
6947 // Get the cost of the scalar memory instruction and address computation.
6948 unsigned Cost = VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
6949
6950 Cost += VF *
6951 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment,
6952 AS);
6953
6954 // Get the overhead of the extractelement and insertelement instructions
6955 // we might create due to scalarization.
6956 Cost += getScalarizationOverhead(I, VF, TTI);
6957
6958 // If we have a predicated store, it may not be executed for each vector
6959 // lane. Scale the cost by the probability of executing the predicated
6960 // block.
6961 if (Legal->isScalarWithPredication(I))
6962 Cost /= getReciprocalPredBlockProb();
6963
6964 return Cost;
6965}
6966
6967unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
6968 unsigned VF) {
6969 Type *ValTy = getMemInstValueType(I);
6970 Type *VectorTy = ToVectorTy(ValTy, VF);
6971 unsigned Alignment = getMemInstAlignment(I);
6972 Value *Ptr = getPointerOperand(I);
6973 unsigned AS = getMemInstAddressSpace(I);
6974 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6975
6976 assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&
"Stride should be 1 or -1 for consecutive memory access") ? static_cast
<void> (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6977, __PRETTY_FUNCTION__))
6977 "Stride should be 1 or -1 for consecutive memory access")(((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&
"Stride should be 1 or -1 for consecutive memory access") ? static_cast
<void> (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6977, __PRETTY_FUNCTION__))
;
6978 unsigned Cost = 0;
6979 if (Legal->isMaskRequired(I))
6980 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6981 else
6982 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6983
6984 bool Reverse = ConsecutiveStride < 0;
6985 if (Reverse)
6986 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
6987 return Cost;
6988}
6989
6990unsigned LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I,
6991 unsigned VF) {
6992 LoadInst *LI = cast<LoadInst>(I);
6993 Type *ValTy = LI->getType();
6994 Type *VectorTy = ToVectorTy(ValTy, VF);
6995 unsigned Alignment = LI->getAlignment();
6996 unsigned AS = LI->getPointerAddressSpace();
6997
6998 return TTI.getAddressComputationCost(ValTy) +
6999 TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS) +
7000 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy);
7001}
7002
7003unsigned LoopVectorizationCostModel::getGatherScatterCost(Instruction *I,
7004 unsigned VF) {
7005 Type *ValTy = getMemInstValueType(I);
7006 Type *VectorTy = ToVectorTy(ValTy, VF);
7007 unsigned Alignment = getMemInstAlignment(I);
7008 Value *Ptr = getPointerOperand(I);
7009
7010 return TTI.getAddressComputationCost(VectorTy) +
7011 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
7012 Legal->isMaskRequired(I), Alignment);
7013}
7014
7015unsigned LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I,
7016 unsigned VF) {
7017 Type *ValTy = getMemInstValueType(I);
7018 Type *VectorTy = ToVectorTy(ValTy, VF);
7019 unsigned AS = getMemInstAddressSpace(I);
7020
7021 auto Group = Legal->getInterleavedAccessGroup(I);
7022 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7022, __PRETTY_FUNCTION__))
;
7023
7024 unsigned InterleaveFactor = Group->getFactor();
7025 Type *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor);
7026
7027 // Holds the indices of existing members in an interleaved load group.
7028 // An interleaved store group doesn't need this as it doesn't allow gaps.
7029 SmallVector<unsigned, 4> Indices;
7030 if (isa<LoadInst>(I)) {
7031 for (unsigned i = 0; i < InterleaveFactor; i++)
7032 if (Group->getMember(i))
7033 Indices.push_back(i);
7034 }
7035
7036 // Calculate the cost of the whole interleaved group.
7037 unsigned Cost = TTI.getInterleavedMemoryOpCost(I->getOpcode(), WideVecTy,
7038 Group->getFactor(), Indices,
7039 Group->getAlignment(), AS);
7040
7041 if (Group->isReverse())
7042 Cost += Group->getNumMembers() *
7043 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7044 return Cost;
7045}
7046
7047unsigned LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I,
7048 unsigned VF) {
7049
7050 // Calculate scalar cost only. Vectorization cost should be ready at this
7051 // moment.
7052 if (VF == 1) {
7053 Type *ValTy = getMemInstValueType(I);
7054 unsigned Alignment = getMemInstAlignment(I);
7055 unsigned AS = getMemInstAlignment(I);
7056
7057 return TTI.getAddressComputationCost(ValTy) +
7058 TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS);
7059 }
7060 return getWideningCost(I, VF);
7061}
7062
7063LoopVectorizationCostModel::VectorizationCostTy
7064LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
7065 // If we know that this instruction will remain uniform, check the cost of
7066 // the scalar version.
7067 if (isUniformAfterVectorization(I, VF))
7068 VF = 1;
7069
7070 if (VF > 1 && isProfitableToScalarize(I, VF))
7071 return VectorizationCostTy(InstsToScalarize[VF][I], false);
7072
7073 Type *VectorTy;
7074 unsigned C = getInstructionCost(I, VF, VectorTy);
7075
7076 bool TypeNotScalarized =
7077 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
7078 return VectorizationCostTy(C, TypeNotScalarized);
7079}
7080
7081void LoopVectorizationCostModel::setCostBasedWideningDecision(unsigned VF) {
7082 if (VF == 1)
7083 return;
7084 for (BasicBlock *BB : TheLoop->blocks()) {
7085 // For each instruction in the old loop.
7086 for (Instruction &I : *BB) {
7087 Value *Ptr = getPointerOperand(&I);
7088 if (!Ptr)
7089 continue;
7090
7091 if (isa<LoadInst>(&I) && Legal->isUniform(Ptr)) {
7092 // Scalar load + broadcast
7093 unsigned Cost = getUniformMemOpCost(&I, VF);
7094 setWideningDecision(&I, VF, CM_Scalarize, Cost);
7095 continue;
7096 }
7097
7098 // We assume that widening is the best solution when possible.
7099 if (Legal->memoryInstructionCanBeWidened(&I, VF)) {
7100 unsigned Cost = getConsecutiveMemOpCost(&I, VF);
7101 setWideningDecision(&I, VF, CM_Widen, Cost);
7102 continue;
7103 }
7104
7105 // Choose between Interleaving, Gather/Scatter or Scalarization.
7106 unsigned InterleaveCost = UINT_MAX(2147483647 *2U +1U);
7107 unsigned NumAccesses = 1;
7108 if (Legal->isAccessInterleaved(&I)) {
7109 auto Group = Legal->getInterleavedAccessGroup(&I);
7110 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-5.0~svn298304/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7110, __PRETTY_FUNCTION__))
;
7111
7112 // Make one decision for the whole group.
7113 if (getWideningDecision(&I, VF) != CM_Unknown)
7114 continue;
7115
7116 NumAccesses = Group->getNumMembers();
7117 InterleaveCost = getInterleaveGroupCost(&I, VF);
7118 }
7119
7120 unsigned GatherScatterCost =
7121 Legal->isLegalGatherOrScatter(&I)
7122 ? getGatherScatterCost(&I, VF) * NumAccesses
7123 : UINT_MAX(2147483647 *2U +1U);
7124
7125 unsigned ScalarizationCost =
7126 getMemInstScalarizationCost(&I, VF) * NumAccesses;
7127
7128 // Choose better solution for the current VF,
7129 // write down this decision and use it during vectorization.
7130 unsigned Cost;
7131 InstWidening Decision;
7132 if (InterleaveCost <= GatherScatterCost &&
7133 InterleaveCost < ScalarizationCost) {
7134 Decision = CM_Interleave;
7135 Cost = InterleaveCost;
7136 } else if (GatherScatterCost < ScalarizationCost) {
7137 Decision = CM_GatherScatter;
7138 Cost = GatherScatterCost;
7139 } else {
7140 Decision = CM_Scalarize;
7141 Cost = ScalarizationCost;
7142 }
7143 // If the instructions belongs to an interleave group, the whole group
7144 // receives the same decision. The whole group receives the cost, but
7145 // the cost will actually be assigned to one instruction.
7146 if (auto Group = Legal->getInterleavedAccessGroup(&I))
7147 setWideningDecision(Group, VF, Decision, Cost);
7148 else
7149 setWideningDecision(&I, VF, Decision, Cost);
7150 }
7151 }
7152}
7153
7154unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
7155 unsigned VF,
7156 Type *&VectorTy) {
7157 Type *RetTy = I->getType();
7158 if (canTruncateToMinimalBitwidth(I, VF))
7159 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
7160 VectorTy = ToVectorTy(RetTy, VF);
7161 auto SE = PSE.getSE();
7162
7163 // TODO: We need to estimate the cost of intrinsic calls.
7164 switch (I->getOpcode()) {
7165 case Instruction::GetElementPtr:
7166 // We mark this instruction as zero-cost because the cost of GEPs in
7167 // vectorized code depends on whether the corresponding memory instruction
7168 // is scalarized or not. Therefore, we handle GEPs with the memory
7169 // instruction cost.
7170 return 0;
7171 case Instruction::Br: {
7172 return TTI.getCFInstrCost(I->getOpcode());
7173 }
7174 case Instruction::PHI: {
7175 auto *Phi = cast<PHINode>(I);
7176
7177 // First-order recurrences are replaced by vector shuffles inside the loop.
7178 if (VF > 1 && Legal->isFirstOrderRecurrence(Phi))
7179 return TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
7180 VectorTy, VF - 1, VectorTy);
7181
7182 // TODO: IF-converted IFs become selects.
7183 return 0;
7184 }
7185 case Instruction::UDiv:
7186 case Instruction::SDiv:
7187 case Instruction::URem:
7188 case Instruction::SRem:
7189 // If we have a predicated instruction, it may not be executed for each
7190 // vector lane. Get the scalarization cost and scale this amount by the
7191 // probability of executing the predicated block. If the instruction is not
7192 // predicated, we fall through to the next case.
7193 if (VF > 1 && Legal->isScalarWithPredication(I)) {
7194 unsigned Cost = 0;
7195
7196 // These instructions have a non-void type, so account for the phi nodes
7197 // that we will create. This cost is likely to be zero. The phi node
7198 // cost, if any, should be scaled by the block probability because it
7199 // models a copy at the end of each predicated block.
7200 Cost += VF * TTI.getCFInstrCost(Instruction::PHI);
7201
7202 // The cost of the non-predicated instruction.
7203 Cost += VF * TTI.getArithmeticInstrCost(I->getOpcode(), RetTy);
7204
7205 // The cost of insertelement and extractelement instructions needed for
7206 // scalarization.
7207 Cost += getScalarizationOverhead(I, VF, TTI);
7208
7209 // Scale the cost by the probability of executing the predicated blocks.
7210 // This assumes the predicated block for each vector lane is equally
7211 // likely.
7212 return Cost / getReciprocalPredBlockProb();
7213 }
7214 case Instruction::Add:
7215 case Instruction::FAdd:
7216 case Instruction::Sub:
7217 case Instruction::FSub:
7218 case Instruction::Mul:
7219 case Instruction::FMul:
7220 case Instruction::FDiv:
7221 case Instruction::FRem:
7222 case Instruction::Shl:
7223 case Instruction::LShr:
7224 case Instruction::AShr:
7225 case Instruction::And:
7226 case Instruction::Or:
7227 case Instruction::Xor: {
7228 // Since we will replace the stride by 1 the multiplication should go away.
7229 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
7230 return 0;
7231 // Certain instructions can be cheaper to vectorize if they have a constant
7232 // second vector operand. One example of this are shifts on x86.
7233 TargetTransformInfo::OperandValueKind Op1VK =
7234 TargetTransformInfo::OK_AnyValue;
7235 TargetTransformInfo::OperandValueKind Op2VK =
7236 TargetTransformInfo::OK_AnyValue;
7237 TargetTransformInfo::OperandValueProperties Op1VP =
7238 TargetTransformInfo::OP_None;
7239 TargetTransformInfo::OperandValueProperties Op2VP =
7240 TargetTransformInfo::OP_None;
7241 Value *Op2 = I->getOperand(1);
7242
7243 // Check for a splat or for a non uniform vector of constants.
7244 if (isa<ConstantInt>(Op2)) {
7245 ConstantInt *CInt = cast<ConstantInt>(Op2);
7246 if (CInt && CInt->getValue().isPowerOf2())
7247 Op2VP = TargetTransformInfo::OP_PowerOf2;
7248 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
7249 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
7250 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
7251 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
7252 if (SplatValue) {
7253 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
7254 if (CInt && CInt->getValue().isPowerOf2())
7255 Op2VP = TargetTransformInfo::OP_PowerOf2;
7256 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
7257 }
7258 } else if (Legal->isUniform(Op2)) {
7259 Op2VK = TargetTransformInfo::OK_UniformValue;
7260 }
7261 SmallVector<const Value *, 4> Operands(I->operand_values());
7262 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK,
7263 Op2VK, Op1VP, Op2VP, Operands);
7264 }
7265 case Instruction::Select: {
7266 SelectInst *SI = cast<SelectInst>(I);
7267 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
7268 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
7269 Type *CondTy = SI->getCondition()->getType();
7270 if (!ScalarCond)
7271 CondTy = VectorType::get(CondTy, VF);
7272
7273 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
7274 }
7275 case Instruction::ICmp:
7276 case Instruction::FCmp: {
7277 Type *ValTy = I->getOperand(0)->getType();
7278 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
7279 if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF))
7280 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]);
7281 VectorTy = ToVectorTy(ValTy, VF);
7282 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
7283 }
7284 case Instruction::Store:
7285 case Instruction::Load: {
7286 VectorTy = ToVectorTy(getMemInstValueType(I), VF);
7287 return getMemoryInstructionCost(I, VF);
7288 }
7289 case Instruction::ZExt:
7290 case Instruction::SExt:
7291 case Instruction::FPToUI:
7292 case Instruction::FPToSI:
7293 case Instruction::FPExt:
7294 case Instruction::PtrToInt:
7295 case Instruction::IntToPtr:
7296 case Instruction::SIToFP:
7297 case Instruction::UIToFP:
7298 case Instruction::Trunc:
7299 case Instruction::FPTrunc:
7300 case Instruction::BitCast: {
7301 // We optimize the truncation of induction variables having constant
7302 // integer steps. The cost of these truncations is the same as the scalar
7303 // operation.
7304 if (isOptimizableIVTruncate(I, VF)) {
7305 auto *Trunc = cast<TruncInst>(I);
7306 return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(),
7307 Trunc->getSrcTy());
7308 }
7309
7310 Type *SrcScalarTy = I->getOperand(0)->getType();
7311 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
7312 if (canTruncateToMinimalBitwidth(I, VF)) {
7313 // This cast is going to be shrunk. This may remove the cast or it might
7314 // turn it into slightly different cast. For example, if MinBW == 16,
7315 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
7316 //
7317 // Calculate the modified src and dest types.
7318 Type *MinVecTy = VectorTy;</