Bug Summary

File:lib/Transforms/Vectorize/LoopVectorize.cpp
Warning:line 6050, 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/SCCIterator.h"
54#include "llvm/ADT/SetVector.h"
55#include "llvm/ADT/SmallPtrSet.h"
56#include "llvm/ADT/SmallSet.h"
57#include "llvm/ADT/SmallVector.h"
58#include "llvm/ADT/Statistic.h"
59#include "llvm/ADT/StringExtras.h"
60#include "llvm/Analysis/CodeMetrics.h"
61#include "llvm/Analysis/GlobalsModRef.h"
62#include "llvm/Analysis/LoopInfo.h"
63#include "llvm/Analysis/LoopIterator.h"
64#include "llvm/Analysis/LoopPass.h"
65#include "llvm/Analysis/ScalarEvolutionExpander.h"
66#include "llvm/Analysis/ScalarEvolutionExpressions.h"
67#include "llvm/Analysis/ValueTracking.h"
68#include "llvm/Analysis/VectorUtils.h"
69#include "llvm/IR/Constants.h"
70#include "llvm/IR/DataLayout.h"
71#include "llvm/IR/DebugInfo.h"
72#include "llvm/IR/DerivedTypes.h"
73#include "llvm/IR/DiagnosticInfo.h"
74#include "llvm/IR/Dominators.h"
75#include "llvm/IR/Function.h"
76#include "llvm/IR/IRBuilder.h"
77#include "llvm/IR/Instructions.h"
78#include "llvm/IR/IntrinsicInst.h"
79#include "llvm/IR/LLVMContext.h"
80#include "llvm/IR/Module.h"
81#include "llvm/IR/PatternMatch.h"
82#include "llvm/IR/Type.h"
83#include "llvm/IR/User.h"
84#include "llvm/IR/Value.h"
85#include "llvm/IR/ValueHandle.h"
86#include "llvm/IR/Verifier.h"
87#include "llvm/Pass.h"
88#include "llvm/Support/BranchProbability.h"
89#include "llvm/Support/CommandLine.h"
90#include "llvm/Support/Debug.h"
91#include "llvm/Support/raw_ostream.h"
92#include "llvm/Transforms/Scalar.h"
93#include "llvm/Transforms/Utils/BasicBlockUtils.h"
94#include "llvm/Transforms/Utils/Local.h"
95#include "llvm/Transforms/Utils/LoopSimplify.h"
96#include "llvm/Transforms/Utils/LoopUtils.h"
97#include "llvm/Transforms/Utils/LoopVersioning.h"
98#include "llvm/Transforms/Vectorize.h"
99#include <algorithm>
100#include <map>
101#include <tuple>
102
103using namespace llvm;
104using namespace llvm::PatternMatch;
105
106#define LV_NAME"loop-vectorize" "loop-vectorize"
107#define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize"
108
109STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized"
, "Number of loops vectorized", {0}, false}
;
110STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed"
, "Number of loops analyzed for vectorization", {0}, false}
;
111
112static cl::opt<bool>
113 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
114 cl::desc("Enable if-conversion during vectorization."));
115
116/// We don't vectorize loops with a known constant trip count below this number.
117static cl::opt<unsigned> TinyTripCountVectorThreshold(
118 "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
119 cl::desc("Don't vectorize loops with a constant "
120 "trip count that is smaller than this "
121 "value."));
122
123static cl::opt<bool> MaximizeBandwidth(
124 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
125 cl::desc("Maximize bandwidth when selecting vectorization factor which "
126 "will be determined by the smallest type in loop."));
127
128static cl::opt<bool> EnableInterleavedMemAccesses(
129 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
130 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
131
132/// Maximum factor for an interleaved memory access.
133static cl::opt<unsigned> MaxInterleaveGroupFactor(
134 "max-interleave-group-factor", cl::Hidden,
135 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
136 cl::init(8));
137
138/// We don't interleave loops with a known constant trip count below this
139/// number.
140static const unsigned TinyTripCountInterleaveThreshold = 128;
141
142static cl::opt<unsigned> ForceTargetNumScalarRegs(
143 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
144 cl::desc("A flag that overrides the target's number of scalar registers."));
145
146static cl::opt<unsigned> ForceTargetNumVectorRegs(
147 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
148 cl::desc("A flag that overrides the target's number of vector registers."));
149
150/// Maximum vectorization interleave count.
151static const unsigned MaxInterleaveFactor = 16;
152
153static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
154 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
155 cl::desc("A flag that overrides the target's max interleave factor for "
156 "scalar loops."));
157
158static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
159 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's max interleave factor for "
161 "vectorized loops."));
162
163static cl::opt<unsigned> ForceTargetInstructionCost(
164 "force-target-instruction-cost", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's expected cost for "
166 "an instruction to a single constant value. Mostly "
167 "useful for getting consistent testing."));
168
169static cl::opt<unsigned> SmallLoopCost(
170 "small-loop-cost", cl::init(20), cl::Hidden,
171 cl::desc(
172 "The cost of a loop that is considered 'small' by the interleaver."));
173
174static cl::opt<bool> LoopVectorizeWithBlockFrequency(
175 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
176 cl::desc("Enable the use of the block frequency analysis to access PGO "
177 "heuristics minimizing code growth in cold regions and being more "
178 "aggressive in hot regions."));
179
180// Runtime interleave loops for load/store throughput.
181static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
182 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
183 cl::desc(
184 "Enable runtime interleaving until load/store ports are saturated"));
185
186/// The number of stores in a loop that are allowed to need predication.
187static cl::opt<unsigned> NumberOfStoresToPredicate(
188 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
189 cl::desc("Max number of stores to be predicated behind an if."));
190
191static cl::opt<bool> EnableIndVarRegisterHeur(
192 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
193 cl::desc("Count the induction variable only once when interleaving"));
194
195static cl::opt<bool> EnableCondStoresVectorization(
196 "enable-cond-stores-vec", cl::init(true), cl::Hidden,
197 cl::desc("Enable if predication of stores during vectorization."));
198
199static cl::opt<unsigned> MaxNestedScalarReductionIC(
200 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
201 cl::desc("The maximum interleave count to use when interleaving a scalar "
202 "reduction in a nested loop."));
203
204static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
205 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
206 cl::desc("The maximum allowed number of runtime memory checks with a "
207 "vectorize(enable) pragma."));
208
209static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
210 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
211 cl::desc("The maximum number of SCEV checks allowed."));
212
213static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
214 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
215 cl::desc("The maximum number of SCEV checks allowed with a "
216 "vectorize(enable) pragma"));
217
218/// Create an analysis remark that explains why vectorization failed
219///
220/// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
221/// RemarkName is the identifier for the remark. If \p I is passed it is an
222/// instruction that prevents vectorization. Otherwise \p TheLoop is used for
223/// the location of the remark. \return the remark object that can be
224/// streamed to.
225static OptimizationRemarkAnalysis
226createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
227 Instruction *I = nullptr) {
228 Value *CodeRegion = TheLoop->getHeader();
229 DebugLoc DL = TheLoop->getStartLoc();
230
231 if (I) {
232 CodeRegion = I->getParent();
233 // If there is no debug location attached to the instruction, revert back to
234 // using the loop's.
235 if (I->getDebugLoc())
236 DL = I->getDebugLoc();
237 }
238
239 OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
240 R << "loop not vectorized: ";
241 return R;
242}
243
244namespace {
245
246// Forward declarations.
247class LoopVectorizeHints;
248class LoopVectorizationLegality;
249class LoopVectorizationCostModel;
250class LoopVectorizationRequirements;
251
252/// Returns true if the given loop body has a cycle, excluding the loop
253/// itself.
254static bool hasCyclesInLoopBody(const Loop &L) {
255 if (!L.empty())
256 return true;
257
258 for (const auto &SCC :
259 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
260 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
261 if (SCC.size() > 1) {
262 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)
;
263 DEBUG(L.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { L.dump(); } } while (false)
;
264 return true;
265 }
266 }
267 return false;
268}
269
270/// \brief This modifies LoopAccessReport to initialize message with
271/// loop-vectorizer-specific part.
272class VectorizationReport : public LoopAccessReport {
273public:
274 VectorizationReport(Instruction *I = nullptr)
275 : LoopAccessReport("loop not vectorized: ", I) {}
276
277 /// \brief This allows promotion of the loop-access analysis report into the
278 /// loop-vectorizer report. It modifies the message to add the
279 /// loop-vectorizer-specific part of the message.
280 explicit VectorizationReport(const LoopAccessReport &R)
281 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
282 R.getInstr()) {}
283};
284
285/// A helper function for converting Scalar types to vector types.
286/// If the incoming type is void, we return void. If the VF is 1, we return
287/// the scalar type.
288static Type *ToVectorTy(Type *Scalar, unsigned VF) {
289 if (Scalar->isVoidTy() || VF == 1)
290 return Scalar;
291 return VectorType::get(Scalar, VF);
292}
293
294/// A helper function that returns GEP instruction and knows to skip a
295/// 'bitcast'. The 'bitcast' may be skipped if the source and the destination
296/// pointee types of the 'bitcast' have the same size.
297/// For example:
298/// bitcast double** %var to i64* - can be skipped
299/// bitcast double** %var to i8* - can not
300static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
301
302 if (isa<GetElementPtrInst>(Ptr))
303 return cast<GetElementPtrInst>(Ptr);
304
305 if (isa<BitCastInst>(Ptr) &&
306 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) {
307 Type *BitcastTy = Ptr->getType();
308 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy();
309 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy))
310 return nullptr;
311 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType();
312 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType();
313 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout();
314 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty))
315 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
316 }
317 return nullptr;
318}
319
320// FIXME: The following helper functions have multiple implementations
321// in the project. They can be effectively organized in a common Load/Store
322// utilities unit.
323
324/// A helper function that returns the pointer operand of a load or store
325/// instruction.
326static Value *getPointerOperand(Value *I) {
327 if (auto *LI = dyn_cast<LoadInst>(I))
328 return LI->getPointerOperand();
329 if (auto *SI = dyn_cast<StoreInst>(I))
330 return SI->getPointerOperand();
331 return nullptr;
332}
333
334/// A helper function that returns the type of loaded or stored value.
335static Type *getMemInstValueType(Value *I) {
336 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 337, __PRETTY_FUNCTION__))
337 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 337, __PRETTY_FUNCTION__))
;
338 if (auto *LI = dyn_cast<LoadInst>(I))
339 return LI->getType();
340 return cast<StoreInst>(I)->getValueOperand()->getType();
341}
342
343/// A helper function that returns the alignment of load or store instruction.
344static unsigned getMemInstAlignment(Value *I) {
345 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 346, __PRETTY_FUNCTION__))
346 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 346, __PRETTY_FUNCTION__))
;
347 if (auto *LI = dyn_cast<LoadInst>(I))
348 return LI->getAlignment();
349 return cast<StoreInst>(I)->getAlignment();
350}
351
352/// A helper function that returns the address space of the pointer operand of
353/// load or store instruction.
354static unsigned getMemInstAddressSpace(Value *I) {
355 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 356, __PRETTY_FUNCTION__))
356 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 356, __PRETTY_FUNCTION__))
;
357 if (auto *LI = dyn_cast<LoadInst>(I))
358 return LI->getPointerAddressSpace();
359 return cast<StoreInst>(I)->getPointerAddressSpace();
360}
361
362/// A helper function that returns true if the given type is irregular. The
363/// type is irregular if its allocated size doesn't equal the store size of an
364/// element of the corresponding vector type at the given vectorization factor.
365static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
366
367 // Determine if an array of VF elements of type Ty is "bitcast compatible"
368 // with a <VF x Ty> vector.
369 if (VF > 1) {
370 auto *VectorTy = VectorType::get(Ty, VF);
371 return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
372 }
373
374 // If the vectorization factor is one, we just check if an array of type Ty
375 // requires padding between elements.
376 return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
377}
378
379/// A helper function that returns the reciprocal of the block probability of
380/// predicated blocks. If we return X, we are assuming the predicated block
381/// will execute once for for every X iterations of the loop header.
382///
383/// TODO: We should use actual block probability here, if available. Currently,
384/// we always assume predicated blocks have a 50% chance of executing.
385static unsigned getReciprocalPredBlockProb() { return 2; }
386
387/// InnerLoopVectorizer vectorizes loops which contain only one basic
388/// block to a specified vectorization factor (VF).
389/// This class performs the widening of scalars into vectors, or multiple
390/// scalars. This class also implements the following features:
391/// * It inserts an epilogue loop for handling loops that don't have iteration
392/// counts that are known to be a multiple of the vectorization factor.
393/// * It handles the code generation for reduction variables.
394/// * Scalarization (implementation using scalars) of un-vectorizable
395/// instructions.
396/// InnerLoopVectorizer does not perform any vectorization-legality
397/// checks, and relies on the caller to check for the different legality
398/// aspects. The InnerLoopVectorizer relies on the
399/// LoopVectorizationLegality class to provide information about the induction
400/// and reduction variables that were found to a given vectorization factor.
401class InnerLoopVectorizer {
402public:
403 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
404 LoopInfo *LI, DominatorTree *DT,
405 const TargetLibraryInfo *TLI,
406 const TargetTransformInfo *TTI, AssumptionCache *AC,
407 OptimizationRemarkEmitter *ORE, unsigned VecWidth,
408 unsigned UnrollFactor, LoopVectorizationLegality *LVL,
409 LoopVectorizationCostModel *CM)
410 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
411 AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
412 Builder(PSE.getSE()->getContext()), Induction(nullptr),
413 OldInduction(nullptr), VectorLoopValueMap(UnrollFactor, VecWidth),
414 TripCount(nullptr), VectorTripCount(nullptr), Legal(LVL), Cost(CM),
415 AddedSafetyChecks(false) {}
416
417 // Perform the actual loop widening (vectorization).
418 void vectorize() {
419 // Create a new empty loop. Unlink the old loop and connect the new one.
420 createEmptyLoop();
421 // Widen each instruction in the old loop to a new one in the new loop.
422 vectorizeLoop();
423 }
424
425 // Return true if any runtime check is added.
426 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
427
428 virtual ~InnerLoopVectorizer() {}
429
430protected:
431 /// A small list of PHINodes.
432 typedef SmallVector<PHINode *, 4> PhiVector;
433
434 /// A type for vectorized values in the new loop. Each value from the
435 /// original loop, when vectorized, is represented by UF vector values in the
436 /// new unrolled loop, where UF is the unroll factor.
437 typedef SmallVector<Value *, 2> VectorParts;
438
439 /// A type for scalarized values in the new loop. Each value from the
440 /// original loop, when scalarized, is represented by UF x VF scalar values
441 /// in the new unrolled loop, where UF is the unroll factor and VF is the
442 /// vectorization factor.
443 typedef SmallVector<SmallVector<Value *, 4>, 2> ScalarParts;
444
445 // When we if-convert we need to create edge masks. We have to cache values
446 // so that we don't end up with exponential recursion/IR.
447 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
448 EdgeMaskCache;
449
450 /// Create an empty loop, based on the loop ranges of the old loop.
451 void createEmptyLoop();
452
453 /// Set up the values of the IVs correctly when exiting the vector loop.
454 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
455 Value *CountRoundDown, Value *EndValue,
456 BasicBlock *MiddleBlock);
457
458 /// Create a new induction variable inside L.
459 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
460 Value *Step, Instruction *DL);
461 /// Copy and widen the instructions from the old loop.
462 virtual void vectorizeLoop();
463
464 /// Fix a first-order recurrence. This is the second phase of vectorizing
465 /// this phi node.
466 void fixFirstOrderRecurrence(PHINode *Phi);
467
468 /// \brief The Loop exit block may have single value PHI nodes where the
469 /// incoming value is 'Undef'. While vectorizing we only handled real values
470 /// that were defined inside the loop. Here we fix the 'undef case'.
471 /// See PR14725.
472 void fixLCSSAPHIs();
473
474 /// Iteratively sink the scalarized operands of a predicated instruction into
475 /// the block that was created for it.
476 void sinkScalarOperands(Instruction *PredInst);
477
478 /// Predicate conditional instructions that require predication on their
479 /// respective conditions.
480 void predicateInstructions();
481
482 /// Collect the instructions from the original loop that would be trivially
483 /// dead in the vectorized loop if generated.
484 void collectTriviallyDeadInstructions();
485
486 /// Shrinks vector element sizes to the smallest bitwidth they can be legally
487 /// represented as.
488 void truncateToMinimalBitwidths();
489
490 /// A helper function that computes the predicate of the block BB, assuming
491 /// that the header block of the loop is set to True. It returns the *entry*
492 /// mask for the block BB.
493 VectorParts createBlockInMask(BasicBlock *BB);
494 /// A helper function that computes the predicate of the edge between SRC
495 /// and DST.
496 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
497
498 /// A helper function to vectorize a single BB within the innermost loop.
499 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
500
501 /// Vectorize a single PHINode in a block. This method handles the induction
502 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
503 /// arbitrary length vectors.
504 void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF,
505 PhiVector *PV);
506
507 /// Insert the new loop to the loop hierarchy and pass manager
508 /// and update the analysis passes.
509 void updateAnalysis();
510
511 /// This instruction is un-vectorizable. Implement it as a sequence
512 /// of scalars. If \p IfPredicateInstr is true we need to 'hide' each
513 /// scalarized instruction behind an if block predicated on the control
514 /// dependence of the instruction.
515 virtual void scalarizeInstruction(Instruction *Instr,
516 bool IfPredicateInstr = false);
517
518 /// Vectorize Load and Store instructions,
519 virtual void vectorizeMemoryInstruction(Instruction *Instr);
520
521 /// Create a broadcast instruction. This method generates a broadcast
522 /// instruction (shuffle) for loop invariant values and for the induction
523 /// value. If this is the induction variable then we extend it to N, N+1, ...
524 /// this is needed because each iteration in the loop corresponds to a SIMD
525 /// element.
526 virtual Value *getBroadcastInstrs(Value *V);
527
528 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
529 /// to each vector element of Val. The sequence starts at StartIndex.
530 /// \p Opcode is relevant for FP induction variable.
531 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
532 Instruction::BinaryOps Opcode =
533 Instruction::BinaryOpsEnd);
534
535 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
536 /// variable on which to base the steps, \p Step is the size of the step, and
537 /// \p EntryVal is the value from the original loop that maps to the steps.
538 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
539 /// can be a truncate instruction).
540 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal);
541
542 /// Create a vector induction phi node based on an existing scalar one. \p
543 /// EntryVal is the value from the original loop that maps to the vector phi
544 /// node, and \p Step is the loop-invariant step. If \p EntryVal is a
545 /// truncate instruction, instead of widening the original IV, we widen a
546 /// version of the IV truncated to \p EntryVal's type.
547 void createVectorIntInductionPHI(const InductionDescriptor &II, Value *Step,
548 Instruction *EntryVal);
549
550 /// Widen an integer induction variable \p IV. If \p Trunc is provided, the
551 /// induction variable will first be truncated to the corresponding type.
552 void widenIntInduction(PHINode *IV, TruncInst *Trunc = nullptr);
553
554 /// Returns true if an instruction \p I should be scalarized instead of
555 /// vectorized for the chosen vectorization factor.
556 bool shouldScalarizeInstruction(Instruction *I) const;
557
558 /// Returns true if we should generate a scalar version of \p IV.
559 bool needsScalarInduction(Instruction *IV) const;
560
561 /// Return a constant reference to the VectorParts corresponding to \p V from
562 /// the original loop. If the value has already been vectorized, the
563 /// corresponding vector entry in VectorLoopValueMap is returned. If,
564 /// however, the value has a scalar entry in VectorLoopValueMap, we construct
565 /// new vector values on-demand by inserting the scalar values into vectors
566 /// with an insertelement sequence. If the value has been neither vectorized
567 /// nor scalarized, it must be loop invariant, so we simply broadcast the
568 /// value into vectors.
569 const VectorParts &getVectorValue(Value *V);
570
571 /// Return a value in the new loop corresponding to \p V from the original
572 /// loop at unroll index \p Part and vector index \p Lane. If the value has
573 /// been vectorized but not scalarized, the necessary extractelement
574 /// instruction will be generated.
575 Value *getScalarValue(Value *V, unsigned Part, unsigned Lane);
576
577 /// Try to vectorize the interleaved access group that \p Instr belongs to.
578 void vectorizeInterleaveGroup(Instruction *Instr);
579
580 /// Generate a shuffle sequence that will reverse the vector Vec.
581 virtual Value *reverseVector(Value *Vec);
582
583 /// Returns (and creates if needed) the original loop trip count.
584 Value *getOrCreateTripCount(Loop *NewLoop);
585
586 /// Returns (and creates if needed) the trip count of the widened loop.
587 Value *getOrCreateVectorTripCount(Loop *NewLoop);
588
589 /// Emit a bypass check to see if the trip count would overflow, or we
590 /// wouldn't have enough iterations to execute one vector loop.
591 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
592 /// Emit a bypass check to see if the vector trip count is nonzero.
593 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
594 /// Emit a bypass check to see if all of the SCEV assumptions we've
595 /// had to make are correct.
596 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
597 /// Emit bypass checks to check any memory assumptions we may have made.
598 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
599
600 /// Add additional metadata to \p To that was not present on \p Orig.
601 ///
602 /// Currently this is used to add the noalias annotations based on the
603 /// inserted memchecks. Use this for instructions that are *cloned* into the
604 /// vector loop.
605 void addNewMetadata(Instruction *To, const Instruction *Orig);
606
607 /// Add metadata from one instruction to another.
608 ///
609 /// This includes both the original MDs from \p From and additional ones (\see
610 /// addNewMetadata). Use this for *newly created* instructions in the vector
611 /// loop.
612 void addMetadata(Instruction *To, Instruction *From);
613
614 /// \brief Similar to the previous function but it adds the metadata to a
615 /// vector of instructions.
616 void addMetadata(ArrayRef<Value *> To, Instruction *From);
617
618 /// \brief Set the debug location in the builder using the debug location in
619 /// the instruction.
620 void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr);
621
622 /// This is a helper class for maintaining vectorization state. It's used for
623 /// mapping values from the original loop to their corresponding values in
624 /// the new loop. Two mappings are maintained: one for vectorized values and
625 /// one for scalarized values. Vectorized values are represented with UF
626 /// vector values in the new loop, and scalarized values are represented with
627 /// UF x VF scalar values in the new loop. UF and VF are the unroll and
628 /// vectorization factors, respectively.
629 ///
630 /// Entries can be added to either map with initVector and initScalar, which
631 /// initialize and return a constant reference to the new entry. If a
632 /// non-constant reference to a vector entry is required, getVector can be
633 /// used to retrieve a mutable entry. We currently directly modify the mapped
634 /// values during "fix-up" operations that occur once the first phase of
635 /// widening is complete. These operations include type truncation and the
636 /// second phase of recurrence widening.
637 ///
638 /// Otherwise, entries from either map should be accessed using the
639 /// getVectorValue or getScalarValue functions from InnerLoopVectorizer.
640 /// getVectorValue and getScalarValue coordinate to generate a vector or
641 /// scalar value on-demand if one is not yet available. When vectorizing a
642 /// loop, we visit the definition of an instruction before its uses. When
643 /// visiting the definition, we either vectorize or scalarize the
644 /// instruction, creating an entry for it in the corresponding map. (In some
645 /// cases, such as induction variables, we will create both vector and scalar
646 /// entries.) Then, as we encounter uses of the definition, we derive values
647 /// for each scalar or vector use unless such a value is already available.
648 /// For example, if we scalarize a definition and one of its uses is vector,
649 /// we build the required vector on-demand with an insertelement sequence
650 /// when visiting the use. Otherwise, if the use is scalar, we can use the
651 /// existing scalar definition.
652 struct ValueMap {
653
654 /// Construct an empty map with the given unroll and vectorization factors.
655 ValueMap(unsigned UnrollFactor, unsigned VecWidth)
656 : UF(UnrollFactor), VF(VecWidth) {
657 // The unroll and vectorization factors are only used in asserts builds
658 // to verify map entries are sized appropriately.
659 (void)UF;
660 (void)VF;
661 }
662
663 /// \return True if the map has a vector entry for \p Key.
664 bool hasVector(Value *Key) const { return VectorMapStorage.count(Key); }
665
666 /// \return True if the map has a scalar entry for \p Key.
667 bool hasScalar(Value *Key) const { return ScalarMapStorage.count(Key); }
668
669 /// \brief Map \p Key to the given VectorParts \p Entry, and return a
670 /// constant reference to the new vector map entry. The given key should
671 /// not already be in the map, and the given VectorParts should be
672 /// correctly sized for the current unroll factor.
673 const VectorParts &initVector(Value *Key, const VectorParts &Entry) {
674 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 674, __PRETTY_FUNCTION__))
;
675 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 675, __PRETTY_FUNCTION__))
;
676 VectorMapStorage[Key] = Entry;
677 return VectorMapStorage[Key];
678 }
679
680 /// \brief Map \p Key to the given ScalarParts \p Entry, and return a
681 /// constant reference to the new scalar map entry. The given key should
682 /// not already be in the map, and the given ScalarParts should be
683 /// correctly sized for the current unroll and vectorization factors.
684 const ScalarParts &initScalar(Value *Key, const ScalarParts &Entry) {
685 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 685, __PRETTY_FUNCTION__))
;
686 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
687 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
688 [&](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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
689 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
690 }) &&((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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
691 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
;
692 ScalarMapStorage[Key] = Entry;
693 return ScalarMapStorage[Key];
694 }
695
696 /// \return A reference to the vector map entry corresponding to \p Key.
697 /// The key should already be in the map. This function should only be used
698 /// when it's necessary to update values that have already been vectorized.
699 /// This is the case for "fix-up" operations including type truncation and
700 /// the second phase of recurrence vectorization. If a non-const reference
701 /// isn't required, getVectorValue should be used instead.
702 VectorParts &getVector(Value *Key) {
703 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 703, __PRETTY_FUNCTION__))
;
704 return VectorMapStorage.find(Key)->second;
705 }
706
707 /// Retrieve an entry from the vector or scalar maps. The preferred way to
708 /// access an existing mapped entry is with getVectorValue or
709 /// getScalarValue from InnerLoopVectorizer. Until those functions can be
710 /// moved inside ValueMap, we have to declare them as friends.
711 friend const VectorParts &InnerLoopVectorizer::getVectorValue(Value *V);
712 friend Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
713 unsigned Lane);
714
715 private:
716 /// The unroll factor. Each entry in the vector map contains UF vector
717 /// values.
718 unsigned UF;
719
720 /// The vectorization factor. Each entry in the scalar map contains UF x VF
721 /// scalar values.
722 unsigned VF;
723
724 /// The vector and scalar map storage. We use std::map and not DenseMap
725 /// because insertions to DenseMap invalidate its iterators.
726 std::map<Value *, VectorParts> VectorMapStorage;
727 std::map<Value *, ScalarParts> ScalarMapStorage;
728 };
729
730 /// The original loop.
731 Loop *OrigLoop;
732 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
733 /// dynamic knowledge to simplify SCEV expressions and converts them to a
734 /// more usable form.
735 PredicatedScalarEvolution &PSE;
736 /// Loop Info.
737 LoopInfo *LI;
738 /// Dominator Tree.
739 DominatorTree *DT;
740 /// Alias Analysis.
741 AliasAnalysis *AA;
742 /// Target Library Info.
743 const TargetLibraryInfo *TLI;
744 /// Target Transform Info.
745 const TargetTransformInfo *TTI;
746 /// Assumption Cache.
747 AssumptionCache *AC;
748 /// Interface to emit optimization remarks.
749 OptimizationRemarkEmitter *ORE;
750
751 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
752 /// used.
753 ///
754 /// This is currently only used to add no-alias metadata based on the
755 /// memchecks. The actually versioning is performed manually.
756 std::unique_ptr<LoopVersioning> LVer;
757
758 /// The vectorization SIMD factor to use. Each vector will have this many
759 /// vector elements.
760 unsigned VF;
761
762protected:
763 /// The vectorization unroll factor to use. Each scalar is vectorized to this
764 /// many different vector instructions.
765 unsigned UF;
766
767 /// The builder that we use
768 IRBuilder<> Builder;
769
770 // --- Vectorization state ---
771
772 /// The vector-loop preheader.
773 BasicBlock *LoopVectorPreHeader;
774 /// The scalar-loop preheader.
775 BasicBlock *LoopScalarPreHeader;
776 /// Middle Block between the vector and the scalar.
777 BasicBlock *LoopMiddleBlock;
778 /// The ExitBlock of the scalar loop.
779 BasicBlock *LoopExitBlock;
780 /// The vector loop body.
781 BasicBlock *LoopVectorBody;
782 /// The scalar loop body.
783 BasicBlock *LoopScalarBody;
784 /// A list of all bypass blocks. The first block is the entry of the loop.
785 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
786
787 /// The new Induction variable which was added to the new block.
788 PHINode *Induction;
789 /// The induction variable of the old basic block.
790 PHINode *OldInduction;
791
792 /// Maps values from the original loop to their corresponding values in the
793 /// vectorized loop. A key value can map to either vector values, scalar
794 /// values or both kinds of values, depending on whether the key was
795 /// vectorized and scalarized.
796 ValueMap VectorLoopValueMap;
797
798 /// Store instructions that should be predicated, as a pair
799 /// <StoreInst, Predicate>
800 SmallVector<std::pair<Instruction *, Value *>, 4> PredicatedInstructions;
801 EdgeMaskCache MaskCache;
802 /// Trip count of the original loop.
803 Value *TripCount;
804 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
805 Value *VectorTripCount;
806
807 /// The legality analysis.
808 LoopVectorizationLegality *Legal;
809
810 /// The profitablity analysis.
811 LoopVectorizationCostModel *Cost;
812
813 // Record whether runtime checks are added.
814 bool AddedSafetyChecks;
815
816 // Holds instructions from the original loop whose counterparts in the
817 // vectorized loop would be trivially dead if generated. For example,
818 // original induction update instructions can become dead because we
819 // separately emit induction "steps" when generating code for the new loop.
820 // Similarly, we create a new latch condition when setting up the structure
821 // of the new loop, so the old one can become dead.
822 SmallPtrSet<Instruction *, 4> DeadInstructions;
823
824 // Holds the end values for each induction variable. We save the end values
825 // so we can later fix-up the external users of the induction variables.
826 DenseMap<PHINode *, Value *> IVEndValues;
827};
828
829class InnerLoopUnroller : public InnerLoopVectorizer {
830public:
831 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
832 LoopInfo *LI, DominatorTree *DT,
833 const TargetLibraryInfo *TLI,
834 const TargetTransformInfo *TTI, AssumptionCache *AC,
835 OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
836 LoopVectorizationLegality *LVL,
837 LoopVectorizationCostModel *CM)
838 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
839 UnrollFactor, LVL, CM) {}
840
841private:
842 void scalarizeInstruction(Instruction *Instr,
843 bool IfPredicateInstr = false) override;
844 void vectorizeMemoryInstruction(Instruction *Instr) override;
845 Value *getBroadcastInstrs(Value *V) override;
846 Value *getStepVector(Value *Val, int StartIdx, Value *Step,
847 Instruction::BinaryOps Opcode =
848 Instruction::BinaryOpsEnd) override;
849 Value *reverseVector(Value *Vec) override;
850};
851
852/// \brief Look for a meaningful debug location on the instruction or it's
853/// operands.
854static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
855 if (!I)
856 return I;
857
858 DebugLoc Empty;
859 if (I->getDebugLoc() != Empty)
860 return I;
861
862 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
863 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
864 if (OpInst->getDebugLoc() != Empty)
865 return OpInst;
866 }
867
868 return I;
869}
870
871void InnerLoopVectorizer::setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
872 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) {
873 const DILocation *DIL = Inst->getDebugLoc();
874 if (DIL && Inst->getFunction()->isDebugInfoForProfiling())
875 B.SetCurrentDebugLocation(DIL->cloneWithDuplicationFactor(UF * VF));
876 else
877 B.SetCurrentDebugLocation(DIL);
878 } else
879 B.SetCurrentDebugLocation(DebugLoc());
880}
881
882#ifndef NDEBUG
883/// \return string containing a file name and a line # for the given loop.
884static std::string getDebugLocString(const Loop *L) {
885 std::string Result;
886 if (L) {
887 raw_string_ostream OS(Result);
888 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
889 LoopDbgLoc.print(OS);
890 else
891 // Just print the module name.
892 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
893 OS.flush();
894 }
895 return Result;
896}
897#endif
898
899void InnerLoopVectorizer::addNewMetadata(Instruction *To,
900 const Instruction *Orig) {
901 // If the loop was versioned with memchecks, add the corresponding no-alias
902 // metadata.
903 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
904 LVer->annotateInstWithNoAlias(To, Orig);
905}
906
907void InnerLoopVectorizer::addMetadata(Instruction *To,
908 Instruction *From) {
909 propagateMetadata(To, From);
910 addNewMetadata(To, From);
911}
912
913void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
914 Instruction *From) {
915 for (Value *V : To) {
916 if (Instruction *I = dyn_cast<Instruction>(V))
917 addMetadata(I, From);
918 }
919}
920
921/// \brief The group of interleaved loads/stores sharing the same stride and
922/// close to each other.
923///
924/// Each member in this group has an index starting from 0, and the largest
925/// index should be less than interleaved factor, which is equal to the absolute
926/// value of the access's stride.
927///
928/// E.g. An interleaved load group of factor 4:
929/// for (unsigned i = 0; i < 1024; i+=4) {
930/// a = A[i]; // Member of index 0
931/// b = A[i+1]; // Member of index 1
932/// d = A[i+3]; // Member of index 3
933/// ...
934/// }
935///
936/// An interleaved store group of factor 4:
937/// for (unsigned i = 0; i < 1024; i+=4) {
938/// ...
939/// A[i] = a; // Member of index 0
940/// A[i+1] = b; // Member of index 1
941/// A[i+2] = c; // Member of index 2
942/// A[i+3] = d; // Member of index 3
943/// }
944///
945/// Note: the interleaved load group could have gaps (missing members), but
946/// the interleaved store group doesn't allow gaps.
947class InterleaveGroup {
948public:
949 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
950 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
951 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 951, __PRETTY_FUNCTION__))
;
952
953 Factor = std::abs(Stride);
954 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 954, __PRETTY_FUNCTION__))
;
955
956 Reverse = Stride < 0;
957 Members[0] = Instr;
958 }
959
960 bool isReverse() const { return Reverse; }
961 unsigned getFactor() const { return Factor; }
962 unsigned getAlignment() const { return Align; }
963 unsigned getNumMembers() const { return Members.size(); }
964
965 /// \brief Try to insert a new member \p Instr with index \p Index and
966 /// alignment \p NewAlign. The index is related to the leader and it could be
967 /// negative if it is the new leader.
968 ///
969 /// \returns false if the instruction doesn't belong to the group.
970 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
971 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 971, __PRETTY_FUNCTION__))
;
972
973 int Key = Index + SmallestKey;
974
975 // Skip if there is already a member with the same index.
976 if (Members.count(Key))
977 return false;
978
979 if (Key > LargestKey) {
980 // The largest index is always less than the interleave factor.
981 if (Index >= static_cast<int>(Factor))
982 return false;
983
984 LargestKey = Key;
985 } else if (Key < SmallestKey) {
986 // The largest index is always less than the interleave factor.
987 if (LargestKey - Key >= static_cast<int>(Factor))
988 return false;
989
990 SmallestKey = Key;
991 }
992
993 // It's always safe to select the minimum alignment.
994 Align = std::min(Align, NewAlign);
995 Members[Key] = Instr;
996 return true;
997 }
998
999 /// \brief Get the member with the given index \p Index
1000 ///
1001 /// \returns nullptr if contains no such member.
1002 Instruction *getMember(unsigned Index) const {
1003 int Key = SmallestKey + Index;
1004 if (!Members.count(Key))
1005 return nullptr;
1006
1007 return Members.find(Key)->second;
1008 }
1009
1010 /// \brief Get the index for the given member. Unlike the key in the member
1011 /// map, the index starts from 0.
1012 unsigned getIndex(Instruction *Instr) const {
1013 for (auto I : Members)
1014 if (I.second == Instr)
1015 return I.first - SmallestKey;
1016
1017 llvm_unreachable("InterleaveGroup contains no such member")::llvm::llvm_unreachable_internal("InterleaveGroup contains no such member"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1017)
;
1018 }
1019
1020 Instruction *getInsertPos() const { return InsertPos; }
1021 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
1022
1023private:
1024 unsigned Factor; // Interleave Factor.
1025 bool Reverse;
1026 unsigned Align;
1027 DenseMap<int, Instruction *> Members;
1028 int SmallestKey;
1029 int LargestKey;
1030
1031 // To avoid breaking dependences, vectorized instructions of an interleave
1032 // group should be inserted at either the first load or the last store in
1033 // program order.
1034 //
1035 // E.g. %even = load i32 // Insert Position
1036 // %add = add i32 %even // Use of %even
1037 // %odd = load i32
1038 //
1039 // store i32 %even
1040 // %odd = add i32 // Def of %odd
1041 // store i32 %odd // Insert Position
1042 Instruction *InsertPos;
1043};
1044
1045/// \brief Drive the analysis of interleaved memory accesses in the loop.
1046///
1047/// Use this class to analyze interleaved accesses only when we can vectorize
1048/// a loop. Otherwise it's meaningless to do analysis as the vectorization
1049/// on interleaved accesses is unsafe.
1050///
1051/// The analysis collects interleave groups and records the relationships
1052/// between the member and the group in a map.
1053class InterleavedAccessInfo {
1054public:
1055 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1056 DominatorTree *DT, LoopInfo *LI)
1057 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
1058 RequiresScalarEpilogue(false) {}
1059
1060 ~InterleavedAccessInfo() {
1061 SmallSet<InterleaveGroup *, 4> DelSet;
1062 // Avoid releasing a pointer twice.
1063 for (auto &I : InterleaveGroupMap)
1064 DelSet.insert(I.second);
1065 for (auto *Ptr : DelSet)
1066 delete Ptr;
1067 }
1068
1069 /// \brief Analyze the interleaved accesses and collect them in interleave
1070 /// groups. Substitute symbolic strides using \p Strides.
1071 void analyzeInterleaving(const ValueToValueMap &Strides);
1072
1073 /// \brief Check if \p Instr belongs to any interleave group.
1074 bool isInterleaved(Instruction *Instr) const {
1075 return InterleaveGroupMap.count(Instr);
1076 }
1077
1078 /// \brief Return the maximum interleave factor of all interleaved groups.
1079 unsigned getMaxInterleaveFactor() const {
1080 unsigned MaxFactor = 1;
1081 for (auto &Entry : InterleaveGroupMap)
1082 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
1083 return MaxFactor;
1084 }
1085
1086 /// \brief Get the interleave group that \p Instr belongs to.
1087 ///
1088 /// \returns nullptr if doesn't have such group.
1089 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1090 if (InterleaveGroupMap.count(Instr))
1091 return InterleaveGroupMap.find(Instr)->second;
1092 return nullptr;
1093 }
1094
1095 /// \brief Returns true if an interleaved group that may access memory
1096 /// out-of-bounds requires a scalar epilogue iteration for correctness.
1097 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1098
1099 /// \brief Initialize the LoopAccessInfo used for dependence checking.
1100 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1101
1102private:
1103 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1104 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1105 /// The interleaved access analysis can also add new predicates (for example
1106 /// by versioning strides of pointers).
1107 PredicatedScalarEvolution &PSE;
1108 Loop *TheLoop;
1109 DominatorTree *DT;
1110 LoopInfo *LI;
1111 const LoopAccessInfo *LAI;
1112
1113 /// True if the loop may contain non-reversed interleaved groups with
1114 /// out-of-bounds accesses. We ensure we don't speculatively access memory
1115 /// out-of-bounds by executing at least one scalar epilogue iteration.
1116 bool RequiresScalarEpilogue;
1117
1118 /// Holds the relationships between the members and the interleave group.
1119 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1120
1121 /// Holds dependences among the memory accesses in the loop. It maps a source
1122 /// access to a set of dependent sink accesses.
1123 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
1124
1125 /// \brief The descriptor for a strided memory access.
1126 struct StrideDescriptor {
1127 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1128 unsigned Align)
1129 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1130
1131 StrideDescriptor() = default;
1132
1133 // The access's stride. It is negative for a reverse access.
1134 int64_t Stride = 0;
1135 const SCEV *Scev = nullptr; // The scalar expression of this access
1136 uint64_t Size = 0; // The size of the memory object.
1137 unsigned Align = 0; // The alignment of this access.
1138 };
1139
1140 /// \brief A type for holding instructions and their stride descriptors.
1141 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
1142
1143 /// \brief Create a new interleave group with the given instruction \p Instr,
1144 /// stride \p Stride and alignment \p Align.
1145 ///
1146 /// \returns the newly created interleave group.
1147 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1148 unsigned Align) {
1149 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1150, __PRETTY_FUNCTION__))
1150 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1150, __PRETTY_FUNCTION__))
;
1151 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1152 return InterleaveGroupMap[Instr];
1153 }
1154
1155 /// \brief Release the group and remove all the relationships.
1156 void releaseGroup(InterleaveGroup *Group) {
1157 for (unsigned i = 0; i < Group->getFactor(); i++)
1158 if (Instruction *Member = Group->getMember(i))
1159 InterleaveGroupMap.erase(Member);
1160
1161 delete Group;
1162 }
1163
1164 /// \brief Collect all the accesses with a constant stride in program order.
1165 void collectConstStrideAccesses(
1166 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1167 const ValueToValueMap &Strides);
1168
1169 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1170 static bool isStrided(int Stride) {
1171 unsigned Factor = std::abs(Stride);
1172 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1173 }
1174
1175 /// \brief Returns true if \p BB is a predicated block.
1176 bool isPredicated(BasicBlock *BB) const {
1177 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1178 }
1179
1180 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1181 bool areDependencesValid() const {
1182 return LAI && LAI->getDepChecker().getDependences();
1183 }
1184
1185 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1186 /// necessary, when constructing interleaved groups.
1187 ///
1188 /// \p A must precede \p B in program order. We return false if reordering is
1189 /// not necessary or is prevented because \p A and \p B may be dependent.
1190 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1191 StrideEntry *B) const {
1192
1193 // Code motion for interleaved accesses can potentially hoist strided loads
1194 // and sink strided stores. The code below checks the legality of the
1195 // following two conditions:
1196 //
1197 // 1. Potentially moving a strided load (B) before any store (A) that
1198 // precedes B, or
1199 //
1200 // 2. Potentially moving a strided store (A) after any load or store (B)
1201 // that A precedes.
1202 //
1203 // It's legal to reorder A and B if we know there isn't a dependence from A
1204 // to B. Note that this determination is conservative since some
1205 // dependences could potentially be reordered safely.
1206
1207 // A is potentially the source of a dependence.
1208 auto *Src = A->first;
1209 auto SrcDes = A->second;
1210
1211 // B is potentially the sink of a dependence.
1212 auto *Sink = B->first;
1213 auto SinkDes = B->second;
1214
1215 // Code motion for interleaved accesses can't violate WAR dependences.
1216 // Thus, reordering is legal if the source isn't a write.
1217 if (!Src->mayWriteToMemory())
1218 return true;
1219
1220 // At least one of the accesses must be strided.
1221 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1222 return true;
1223
1224 // If dependence information is not available from LoopAccessInfo,
1225 // conservatively assume the instructions can't be reordered.
1226 if (!areDependencesValid())
1227 return false;
1228
1229 // If we know there is a dependence from source to sink, assume the
1230 // instructions can't be reordered. Otherwise, reordering is legal.
1231 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1232 }
1233
1234 /// \brief Collect the dependences from LoopAccessInfo.
1235 ///
1236 /// We process the dependences once during the interleaved access analysis to
1237 /// enable constant-time dependence queries.
1238 void collectDependences() {
1239 if (!areDependencesValid())
1240 return;
1241 auto *Deps = LAI->getDepChecker().getDependences();
1242 for (auto Dep : *Deps)
1243 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1244 }
1245};
1246
1247/// Utility class for getting and setting loop vectorizer hints in the form
1248/// of loop metadata.
1249/// This class keeps a number of loop annotations locally (as member variables)
1250/// and can, upon request, write them back as metadata on the loop. It will
1251/// initially scan the loop for existing metadata, and will update the local
1252/// values based on information in the loop.
1253/// We cannot write all values to metadata, as the mere presence of some info,
1254/// for example 'force', means a decision has been made. So, we need to be
1255/// careful NOT to add them if the user hasn't specifically asked so.
1256class LoopVectorizeHints {
1257 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1258
1259 /// Hint - associates name and validation with the hint value.
1260 struct Hint {
1261 const char *Name;
1262 unsigned Value; // This may have to change for non-numeric values.
1263 HintKind Kind;
1264
1265 Hint(const char *Name, unsigned Value, HintKind Kind)
1266 : Name(Name), Value(Value), Kind(Kind) {}
1267
1268 bool validate(unsigned Val) {
1269 switch (Kind) {
1270 case HK_WIDTH:
1271 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1272 case HK_UNROLL:
1273 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1274 case HK_FORCE:
1275 return (Val <= 1);
1276 }
1277 return false;
1278 }
1279 };
1280
1281 /// Vectorization width.
1282 Hint Width;
1283 /// Vectorization interleave factor.
1284 Hint Interleave;
1285 /// Vectorization forced
1286 Hint Force;
1287
1288 /// Return the loop metadata prefix.
1289 static StringRef Prefix() { return "llvm.loop."; }
1290
1291 /// True if there is any unsafe math in the loop.
1292 bool PotentiallyUnsafe;
1293
1294public:
1295 enum ForceKind {
1296 FK_Undefined = -1, ///< Not selected.
1297 FK_Disabled = 0, ///< Forcing disabled.
1298 FK_Enabled = 1, ///< Forcing enabled.
1299 };
1300
1301 LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1302 OptimizationRemarkEmitter &ORE)
1303 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1304 HK_WIDTH),
1305 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1306 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1307 PotentiallyUnsafe(false), TheLoop(L), ORE(ORE) {
1308 // Populate values with existing loop metadata.
1309 getHintsFromMetadata();
1310
1311 // force-vector-interleave overrides DisableInterleaving.
1312 if (VectorizerParams::isInterleaveForced())
1313 Interleave.Value = VectorizerParams::VectorizationInterleave;
1314
1315 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)
1316 << "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)
;
1317 }
1318
1319 /// Mark the loop L as already vectorized by setting the width to 1.
1320 void setAlreadyVectorized() {
1321 Width.Value = Interleave.Value = 1;
1322 Hint Hints[] = {Width, Interleave};
1323 writeHintsToMetadata(Hints);
1324 }
1325
1326 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1327 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1328 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)
;
1329 emitRemarkWithHints();
1330 return false;
1331 }
1332
1333 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1334 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)
;
1335 emitRemarkWithHints();
1336 return false;
1337 }
1338
1339 if (getWidth() == 1 && getInterleave() == 1) {
1340 // FIXME: Add a separate metadata to indicate when the loop has already
1341 // been vectorized instead of setting width and count to 1.
1342 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)
;
1343 // FIXME: Add interleave.disable metadata. This will allow
1344 // vectorize.disable to be used without disabling the pass and errors
1345 // to differentiate between disabled vectorization and a width of 1.
1346 ORE.emit(OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1347 "AllDisabled", L->getStartLoc(),
1348 L->getHeader())
1349 << "loop not vectorized: vectorization and interleaving are "
1350 "explicitly disabled, or vectorize width and interleave "
1351 "count are both set to 1");
1352 return false;
1353 }
1354
1355 return true;
1356 }
1357
1358 /// Dumps all the hint information.
1359 void emitRemarkWithHints() const {
1360 using namespace ore;
1361 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1362 ORE.emit(OptimizationRemarkMissed(LV_NAME"loop-vectorize", "MissedExplicitlyDisabled",
1363 TheLoop->getStartLoc(),
1364 TheLoop->getHeader())
1365 << "loop not vectorized: vectorization is explicitly disabled");
1366 else {
1367 OptimizationRemarkMissed R(LV_NAME"loop-vectorize", "MissedDetails",
1368 TheLoop->getStartLoc(), TheLoop->getHeader());
1369 R << "loop not vectorized";
1370 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1371 R << " (Force=" << NV("Force", true);
1372 if (Width.Value != 0)
1373 R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1374 if (Interleave.Value != 0)
1375 R << ", Interleave Count=" << NV("InterleaveCount", Interleave.Value);
1376 R << ")";
1377 }
1378 ORE.emit(R);
1379 }
1380 }
1381
1382 unsigned getWidth() const { return Width.Value; }
1383 unsigned getInterleave() const { return Interleave.Value; }
1384 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1385
1386 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1387 /// pass name to force the frontend to print the diagnostic.
1388 const char *vectorizeAnalysisPassName() const {
1389 if (getWidth() == 1)
1390 return LV_NAME"loop-vectorize";
1391 if (getForce() == LoopVectorizeHints::FK_Disabled)
1392 return LV_NAME"loop-vectorize";
1393 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1394 return LV_NAME"loop-vectorize";
1395 return OptimizationRemarkAnalysis::AlwaysPrint;
1396 }
1397
1398 bool allowReordering() const {
1399 // When enabling loop hints are provided we allow the vectorizer to change
1400 // the order of operations that is given by the scalar loop. This is not
1401 // enabled by default because can be unsafe or inefficient. For example,
1402 // reordering floating-point operations will change the way round-off
1403 // error accumulates in the loop.
1404 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1405 }
1406
1407 bool isPotentiallyUnsafe() const {
1408 // Avoid FP vectorization if the target is unsure about proper support.
1409 // This may be related to the SIMD unit in the target not handling
1410 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1411 // Otherwise, a sequence of vectorized loops, even without reduction,
1412 // could lead to different end results on the destination vectors.
1413 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1414 }
1415
1416 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1417
1418private:
1419 /// Find hints specified in the loop metadata and update local values.
1420 void getHintsFromMetadata() {
1421 MDNode *LoopID = TheLoop->getLoopID();
1422 if (!LoopID)
1423 return;
1424
1425 // First operand should refer to the loop id itself.
1426 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1426, __PRETTY_FUNCTION__))
;
1427 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1427, __PRETTY_FUNCTION__))
;
1428
1429 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1430 const MDString *S = nullptr;
1431 SmallVector<Metadata *, 4> Args;
1432
1433 // The expected hint is either a MDString or a MDNode with the first
1434 // operand a MDString.
1435 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1436 if (!MD || MD->getNumOperands() == 0)
1437 continue;
1438 S = dyn_cast<MDString>(MD->getOperand(0));
1439 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1440 Args.push_back(MD->getOperand(i));
1441 } else {
1442 S = dyn_cast<MDString>(LoopID->getOperand(i));
1443 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1443, __PRETTY_FUNCTION__))
;
1444 }
1445
1446 if (!S)
1447 continue;
1448
1449 // Check if the hint starts with the loop metadata prefix.
1450 StringRef Name = S->getString();
1451 if (Args.size() == 1)
1452 setHint(Name, Args[0]);
1453 }
1454 }
1455
1456 /// Checks string hint with one operand and set value if valid.
1457 void setHint(StringRef Name, Metadata *Arg) {
1458 if (!Name.startswith(Prefix()))
1459 return;
1460 Name = Name.substr(Prefix().size(), StringRef::npos);
1461
1462 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1463 if (!C)
1464 return;
1465 unsigned Val = C->getZExtValue();
1466
1467 Hint *Hints[] = {&Width, &Interleave, &Force};
1468 for (auto H : Hints) {
1469 if (Name == H->Name) {
1470 if (H->validate(Val))
1471 H->Value = Val;
1472 else
1473 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)
;
1474 break;
1475 }
1476 }
1477 }
1478
1479 /// Create a new hint from name / value pair.
1480 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1481 LLVMContext &Context = TheLoop->getHeader()->getContext();
1482 Metadata *MDs[] = {MDString::get(Context, Name),
1483 ConstantAsMetadata::get(
1484 ConstantInt::get(Type::getInt32Ty(Context), V))};
1485 return MDNode::get(Context, MDs);
1486 }
1487
1488 /// Matches metadata with hint name.
1489 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1490 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1491 if (!Name)
1492 return false;
1493
1494 for (auto H : HintTypes)
1495 if (Name->getString().endswith(H.Name))
1496 return true;
1497 return false;
1498 }
1499
1500 /// Sets current hints into loop metadata, keeping other values intact.
1501 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1502 if (HintTypes.size() == 0)
1503 return;
1504
1505 // Reserve the first element to LoopID (see below).
1506 SmallVector<Metadata *, 4> MDs(1);
1507 // If the loop already has metadata, then ignore the existing operands.
1508 MDNode *LoopID = TheLoop->getLoopID();
1509 if (LoopID) {
1510 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1511 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1512 // If node in update list, ignore old value.
1513 if (!matchesHintMetadataName(Node, HintTypes))
1514 MDs.push_back(Node);
1515 }
1516 }
1517
1518 // Now, add the missing hints.
1519 for (auto H : HintTypes)
1520 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1521
1522 // Replace current metadata node with new one.
1523 LLVMContext &Context = TheLoop->getHeader()->getContext();
1524 MDNode *NewLoopID = MDNode::get(Context, MDs);
1525 // Set operand 0 to refer to the loop id itself.
1526 NewLoopID->replaceOperandWith(0, NewLoopID);
1527
1528 TheLoop->setLoopID(NewLoopID);
1529 }
1530
1531 /// The loop these hints belong to.
1532 const Loop *TheLoop;
1533
1534 /// Interface to emit optimization remarks.
1535 OptimizationRemarkEmitter &ORE;
1536};
1537
1538static void emitAnalysisDiag(const Loop *TheLoop,
1539 const LoopVectorizeHints &Hints,
1540 OptimizationRemarkEmitter &ORE,
1541 const LoopAccessReport &Message) {
1542 const char *Name = Hints.vectorizeAnalysisPassName();
1543 LoopAccessReport::emitAnalysis(Message, TheLoop, Name, ORE);
1544}
1545
1546static void emitMissedWarning(Function *F, Loop *L,
1547 const LoopVectorizeHints &LH,
1548 OptimizationRemarkEmitter *ORE) {
1549 LH.emitRemarkWithHints();
1550
1551 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1552 if (LH.getWidth() != 1)
1553 ORE->emit(DiagnosticInfoOptimizationFailure(
1554 DEBUG_TYPE"loop-vectorize", "FailedRequestedVectorization",
1555 L->getStartLoc(), L->getHeader())
1556 << "loop not vectorized: "
1557 << "failed explicitly specified loop vectorization");
1558 else if (LH.getInterleave() != 1)
1559 ORE->emit(DiagnosticInfoOptimizationFailure(
1560 DEBUG_TYPE"loop-vectorize", "FailedRequestedInterleaving", L->getStartLoc(),
1561 L->getHeader())
1562 << "loop not interleaved: "
1563 << "failed explicitly specified loop interleaving");
1564 }
1565}
1566
1567/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1568/// to what vectorization factor.
1569/// This class does not look at the profitability of vectorization, only the
1570/// legality. This class has two main kinds of checks:
1571/// * Memory checks - The code in canVectorizeMemory checks if vectorization
1572/// will change the order of memory accesses in a way that will change the
1573/// correctness of the program.
1574/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1575/// checks for a number of different conditions, such as the availability of a
1576/// single induction variable, that all types are supported and vectorize-able,
1577/// etc. This code reflects the capabilities of InnerLoopVectorizer.
1578/// This class is also used by InnerLoopVectorizer for identifying
1579/// induction variable and the different reduction variables.
1580class LoopVectorizationLegality {
1581public:
1582 LoopVectorizationLegality(
1583 Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
1584 TargetLibraryInfo *TLI, AliasAnalysis *AA, Function *F,
1585 const TargetTransformInfo *TTI,
1586 std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1587 OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R,
1588 LoopVectorizeHints *H)
1589 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TTI(TTI), DT(DT),
1590 GetLAA(GetLAA), LAI(nullptr), ORE(ORE), InterleaveInfo(PSE, L, DT, LI),
1591 PrimaryInduction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1592 Requirements(R), Hints(H) {}
1593
1594 /// ReductionList contains the reduction descriptors for all
1595 /// of the reductions that were found in the loop.
1596 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1597
1598 /// InductionList saves induction variables and maps them to the
1599 /// induction descriptor.
1600 typedef MapVector<PHINode *, InductionDescriptor> InductionList;
1601
1602 /// RecurrenceSet contains the phi nodes that are recurrences other than
1603 /// inductions and reductions.
1604 typedef SmallPtrSet<const PHINode *, 8> RecurrenceSet;
1605
1606 /// Returns true if it is legal to vectorize this loop.
1607 /// This does not mean that it is profitable to vectorize this
1608 /// loop, only that it is legal to do so.
1609 bool canVectorize();
1610
1611 /// Returns the primary induction variable.
1612 PHINode *getPrimaryInduction() { return PrimaryInduction; }
1613
1614 /// Returns the reduction variables found in the loop.
1615 ReductionList *getReductionVars() { return &Reductions; }
1616
1617 /// Returns the induction variables found in the loop.
1618 InductionList *getInductionVars() { return &Inductions; }
1619
1620 /// Return the first-order recurrences found in the loop.
1621 RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1622
1623 /// Returns the widest induction type.
1624 Type *getWidestInductionType() { return WidestIndTy; }
1625
1626 /// Returns True if V is an induction variable in this loop.
1627 bool isInductionVariable(const Value *V);
1628
1629 /// Returns True if PN is a reduction variable in this loop.
1630 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1631
1632 /// Returns True if Phi is a first-order recurrence in this loop.
1633 bool isFirstOrderRecurrence(const PHINode *Phi);
1634
1635 /// Return true if the block BB needs to be predicated in order for the loop
1636 /// to be vectorized.
1637 bool blockNeedsPredication(BasicBlock *BB);
1638
1639 /// Check if this pointer is consecutive when vectorizing. This happens
1640 /// when the last index of the GEP is the induction variable, or that the
1641 /// pointer itself is an induction variable.
1642 /// This check allows us to vectorize A[idx] into a wide load/store.
1643 /// Returns:
1644 /// 0 - Stride is unknown or non-consecutive.
1645 /// 1 - Address is consecutive.
1646 /// -1 - Address is consecutive, and decreasing.
1647 int isConsecutivePtr(Value *Ptr);
1648
1649 /// Returns true if the value V is uniform within the loop.
1650 bool isUniform(Value *V);
1651
1652 /// Returns the information that we collected about runtime memory check.
1653 const RuntimePointerChecking *getRuntimePointerChecking() const {
1654 return LAI->getRuntimePointerChecking();
1655 }
1656
1657 const LoopAccessInfo *getLAI() const { return LAI; }
1658
1659 /// \brief Check if \p Instr belongs to any interleaved access group.
1660 bool isAccessInterleaved(Instruction *Instr) {
1661 return InterleaveInfo.isInterleaved(Instr);
1662 }
1663
1664 /// \brief Return the maximum interleave factor of all interleaved groups.
1665 unsigned getMaxInterleaveFactor() const {
1666 return InterleaveInfo.getMaxInterleaveFactor();
1667 }
1668
1669 /// \brief Get the interleaved access group that \p Instr belongs to.
1670 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1671 return InterleaveInfo.getInterleaveGroup(Instr);
1672 }
1673
1674 /// \brief Returns true if an interleaved group requires a scalar iteration
1675 /// to handle accesses with gaps.
1676 bool requiresScalarEpilogue() const {
1677 return InterleaveInfo.requiresScalarEpilogue();
1678 }
1679
1680 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1681
1682 bool hasStride(Value *V) { return LAI->hasStride(V); }
1683
1684 /// Returns true if the target machine supports masked store operation
1685 /// for the given \p DataType and kind of access to \p Ptr.
1686 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1687 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1688 }
1689 /// Returns true if the target machine supports masked load operation
1690 /// for the given \p DataType and kind of access to \p Ptr.
1691 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1692 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1693 }
1694 /// Returns true if the target machine supports masked scatter operation
1695 /// for the given \p DataType.
1696 bool isLegalMaskedScatter(Type *DataType) {
1697 return TTI->isLegalMaskedScatter(DataType);
1698 }
1699 /// Returns true if the target machine supports masked gather operation
1700 /// for the given \p DataType.
1701 bool isLegalMaskedGather(Type *DataType) {
1702 return TTI->isLegalMaskedGather(DataType);
1703 }
1704 /// Returns true if the target machine can represent \p V as a masked gather
1705 /// or scatter operation.
1706 bool isLegalGatherOrScatter(Value *V) {
1707 auto *LI = dyn_cast<LoadInst>(V);
1708 auto *SI = dyn_cast<StoreInst>(V);
1709 if (!LI && !SI)
1710 return false;
1711 auto *Ptr = getPointerOperand(V);
1712 auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1713 return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1714 }
1715
1716 /// Returns true if vector representation of the instruction \p I
1717 /// requires mask.
1718 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1719 unsigned getNumStores() const { return LAI->getNumStores(); }
1720 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1721 unsigned getNumPredStores() const { return NumPredStores; }
1722
1723 /// Returns true if \p I is an instruction that will be scalarized with
1724 /// predication. Such instructions include conditional stores and
1725 /// instructions that may divide by zero.
1726 bool isScalarWithPredication(Instruction *I);
1727
1728 /// Returns true if \p I is a memory instruction with consecutive memory
1729 /// access that can be widened.
1730 bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
1731
1732private:
1733 /// Check if a single basic block loop is vectorizable.
1734 /// At this point we know that this is a loop with a constant trip count
1735 /// and we only need to check individual instructions.
1736 bool canVectorizeInstrs();
1737
1738 /// When we vectorize loops we may change the order in which
1739 /// we read and write from memory. This method checks if it is
1740 /// legal to vectorize the code, considering only memory constrains.
1741 /// Returns true if the loop is vectorizable
1742 bool canVectorizeMemory();
1743
1744 /// Return true if we can vectorize this loop using the IF-conversion
1745 /// transformation.
1746 bool canVectorizeWithIfConvert();
1747
1748 /// Return true if all of the instructions in the block can be speculatively
1749 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1750 /// and we know that we can read from them without segfault.
1751 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1752
1753 /// Updates the vectorization state by adding \p Phi to the inductions list.
1754 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1755 /// better choice for the main induction than the existing one.
1756 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1757 SmallPtrSetImpl<Value *> &AllowedExit);
1758
1759 /// Report an analysis message to assist the user in diagnosing loops that are
1760 /// not vectorized. These are handled as LoopAccessReport rather than
1761 /// VectorizationReport because the << operator of VectorizationReport returns
1762 /// LoopAccessReport.
1763 void emitAnalysis(const LoopAccessReport &Message) const {
1764 emitAnalysisDiag(TheLoop, *Hints, *ORE, Message);
1765 }
1766
1767 /// Create an analysis remark that explains why vectorization failed
1768 ///
1769 /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1770 /// an instruction that prevents vectorization. Otherwise the loop is used
1771 /// for the location of the remark. \return the remark object that can be
1772 /// streamed to.
1773 OptimizationRemarkAnalysis
1774 createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1775 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1776 RemarkName, TheLoop, I);
1777 }
1778
1779 /// \brief If an access has a symbolic strides, this maps the pointer value to
1780 /// the stride symbol.
1781 const ValueToValueMap *getSymbolicStrides() {
1782 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1783 // it's collected. This happens from canVectorizeWithIfConvert, when the
1784 // pointer is checked to reference consecutive elements suitable for a
1785 // masked access.
1786 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1787 }
1788
1789 unsigned NumPredStores;
1790
1791 /// The loop that we evaluate.
1792 Loop *TheLoop;
1793 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1794 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1795 /// of existing SCEV assumptions. The analysis will also add a minimal set
1796 /// of new predicates if this is required to enable vectorization and
1797 /// unrolling.
1798 PredicatedScalarEvolution &PSE;
1799 /// Target Library Info.
1800 TargetLibraryInfo *TLI;
1801 /// Target Transform Info
1802 const TargetTransformInfo *TTI;
1803 /// Dominator Tree.
1804 DominatorTree *DT;
1805 // LoopAccess analysis.
1806 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1807 // And the loop-accesses info corresponding to this loop. This pointer is
1808 // null until canVectorizeMemory sets it up.
1809 const LoopAccessInfo *LAI;
1810 /// Interface to emit optimization remarks.
1811 OptimizationRemarkEmitter *ORE;
1812
1813 /// The interleave access information contains groups of interleaved accesses
1814 /// with the same stride and close to each other.
1815 InterleavedAccessInfo InterleaveInfo;
1816
1817 // --- vectorization state --- //
1818
1819 /// Holds the primary induction variable. This is the counter of the
1820 /// loop.
1821 PHINode *PrimaryInduction;
1822 /// Holds the reduction variables.
1823 ReductionList Reductions;
1824 /// Holds all of the induction variables that we found in the loop.
1825 /// Notice that inductions don't need to start at zero and that induction
1826 /// variables can be pointers.
1827 InductionList Inductions;
1828 /// Holds the phi nodes that are first-order recurrences.
1829 RecurrenceSet FirstOrderRecurrences;
1830 /// Holds the widest induction type encountered.
1831 Type *WidestIndTy;
1832
1833 /// Allowed outside users. This holds the induction and reduction
1834 /// vars which can be accessed from outside the loop.
1835 SmallPtrSet<Value *, 4> AllowedExit;
1836
1837 /// Can we assume the absence of NaNs.
1838 bool HasFunNoNaNAttr;
1839
1840 /// Vectorization requirements that will go through late-evaluation.
1841 LoopVectorizationRequirements *Requirements;
1842
1843 /// Used to emit an analysis of any legality issues.
1844 LoopVectorizeHints *Hints;
1845
1846 /// While vectorizing these instructions we have to generate a
1847 /// call to the appropriate masked intrinsic
1848 SmallPtrSet<const Instruction *, 8> MaskedOp;
1849};
1850
1851/// LoopVectorizationCostModel - estimates the expected speedups due to
1852/// vectorization.
1853/// In many cases vectorization is not profitable. This can happen because of
1854/// a number of reasons. In this class we mainly attempt to predict the
1855/// expected speedup/slowdowns due to the supported instruction set. We use the
1856/// TargetTransformInfo to query the different backends for the cost of
1857/// different operations.
1858class LoopVectorizationCostModel {
1859public:
1860 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1861 LoopInfo *LI, LoopVectorizationLegality *Legal,
1862 const TargetTransformInfo &TTI,
1863 const TargetLibraryInfo *TLI, DemandedBits *DB,
1864 AssumptionCache *AC,
1865 OptimizationRemarkEmitter *ORE, const Function *F,
1866 const LoopVectorizeHints *Hints)
1867 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1868 AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1869
1870 /// Information about vectorization costs
1871 struct VectorizationFactor {
1872 unsigned Width; // Vector width with best cost
1873 unsigned Cost; // Cost of the loop with that width
1874 };
1875 /// \return The most profitable vectorization factor and the cost of that VF.
1876 /// This method checks every power of two up to VF. If UserVF is not ZERO
1877 /// then this vectorization factor will be selected if vectorization is
1878 /// possible.
1879 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1880
1881 /// \return The size (in bits) of the smallest and widest types in the code
1882 /// that needs to be vectorized. We ignore values that remain scalar such as
1883 /// 64 bit loop indices.
1884 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1885
1886 /// \return The desired interleave count.
1887 /// If interleave count has been specified by metadata it will be returned.
1888 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1889 /// are the selected vectorization factor and the cost of the selected VF.
1890 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1891 unsigned LoopCost);
1892
1893 /// Memory access instruction may be vectorized in more than one way.
1894 /// Form of instruction after vectorization depends on cost.
1895 /// This function takes cost-based decisions for Load/Store instructions
1896 /// and collects them in a map. This decisions map is used for building
1897 /// the lists of loop-uniform and loop-scalar instructions.
1898 /// The calculated cost is saved with widening decision in order to
1899 /// avoid redundant calculations.
1900 void setCostBasedWideningDecision(unsigned VF);
1901
1902 /// \brief A struct that represents some properties of the register usage
1903 /// of a loop.
1904 struct RegisterUsage {
1905 /// Holds the number of loop invariant values that are used in the loop.
1906 unsigned LoopInvariantRegs;
1907 /// Holds the maximum number of concurrent live intervals in the loop.
1908 unsigned MaxLocalUsers;
1909 /// Holds the number of instructions in the loop.
1910 unsigned NumInstructions;
1911 };
1912
1913 /// \return Returns information about the register usages of the loop for the
1914 /// given vectorization factors.
1915 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1916
1917 /// Collect values we want to ignore in the cost model.
1918 void collectValuesToIgnore();
1919
1920 /// \returns The smallest bitwidth each instruction can be represented with.
1921 /// The vector equivalents of these instructions should be truncated to this
1922 /// type.
1923 const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1924 return MinBWs;
1925 }
1926
1927 /// \returns True if it is more profitable to scalarize instruction \p I for
1928 /// vectorization factor \p VF.
1929 bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1930 auto Scalars = InstsToScalarize.find(VF);
1931 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1932, __PRETTY_FUNCTION__))
1932 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1932, __PRETTY_FUNCTION__))
;
1933 return Scalars->second.count(I);
1934 }
1935
1936 /// Returns true if \p I is known to be uniform after vectorization.
1937 bool isUniformAfterVectorization(Instruction *I, unsigned VF) const {
1938 if (VF == 1)
1939 return true;
1940 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1940, __PRETTY_FUNCTION__))
;
1941 auto UniformsPerVF = Uniforms.find(VF);
1942 return UniformsPerVF->second.count(I);
1943 }
1944
1945 /// Returns true if \p I is known to be scalar after vectorization.
1946 bool isScalarAfterVectorization(Instruction *I, unsigned VF) const {
1947 if (VF == 1)
1948 return true;
1949 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1949, __PRETTY_FUNCTION__))
;
1950 auto ScalarsPerVF = Scalars.find(VF);
1951 return ScalarsPerVF->second.count(I);
1952 }
1953
1954 /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1955 /// for vectorization factor \p VF.
1956 bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1957 return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1958 !isScalarAfterVectorization(I, VF);
1959 }
1960
1961 /// Decision that was taken during cost calculation for memory instruction.
1962 enum InstWidening {
1963 CM_Unknown,
1964 CM_Widen,
1965 CM_Interleave,
1966 CM_GatherScatter,
1967 CM_Scalarize
1968 };
1969
1970 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1971 /// instruction \p I and vector width \p VF.
1972 void setWideningDecision(Instruction *I, unsigned VF, InstWidening W,
1973 unsigned Cost) {
1974 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1974, __PRETTY_FUNCTION__))
;
1975 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1976 }
1977
1978 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1979 /// interleaving group \p Grp and vector width \p VF.
1980 void setWideningDecision(const InterleaveGroup *Grp, unsigned VF,
1981 InstWidening W, unsigned Cost) {
1982 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1982, __PRETTY_FUNCTION__))
;
1983 /// Broadcast this decicion to all instructions inside the group.
1984 /// But the cost will be assigned to one instruction only.
1985 for (unsigned i = 0; i < Grp->getFactor(); ++i) {
1986 if (auto *I = Grp->getMember(i)) {
1987 if (Grp->getInsertPos() == I)
1988 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1989 else
1990 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0);
1991 }
1992 }
1993 }
1994
1995 /// Return the cost model decision for the given instruction \p I and vector
1996 /// width \p VF. Return CM_Unknown if this instruction did not pass
1997 /// through the cost modeling.
1998 InstWidening getWideningDecision(Instruction *I, unsigned VF) {
1999 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1999, __PRETTY_FUNCTION__))
;
2000 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
2001 auto Itr = WideningDecisions.find(InstOnVF);
2002 if (Itr == WideningDecisions.end())
2003 return CM_Unknown;
2004 return Itr->second.first;
2005 }
2006
2007 /// Return the vectorization cost for the given instruction \p I and vector
2008 /// width \p VF.
2009 unsigned getWideningCost(Instruction *I, unsigned VF) {
2010 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2010, __PRETTY_FUNCTION__))
;
2011 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
2012 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2012, __PRETTY_FUNCTION__))
;
2013 return WideningDecisions[InstOnVF].second;
2014 }
2015
2016 /// Return True if instruction \p I is an optimizable truncate whose operand
2017 /// is an induction variable. Such a truncate will be removed by adding a new
2018 /// induction variable with the destination type.
2019 bool isOptimizableIVTruncate(Instruction *I, unsigned VF) {
2020
2021 // If the instruction is not a truncate, return false.
2022 auto *Trunc = dyn_cast<TruncInst>(I);
2023 if (!Trunc)
2024 return false;
2025
2026 // Get the source and destination types of the truncate.
2027 Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF);
2028 Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF);
2029
2030 // If the truncate is free for the given types, return false. Replacing a
2031 // free truncate with an induction variable would add an induction variable
2032 // update instruction to each iteration of the loop. We exclude from this
2033 // check the primary induction variable since it will need an update
2034 // instruction regardless.
2035 Value *Op = Trunc->getOperand(0);
2036 if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))
2037 return false;
2038
2039 // If the truncated value is not an induction variable, return false.
2040 return Legal->isInductionVariable(Op);
2041 }
2042
2043private:
2044 /// The vectorization cost is a combination of the cost itself and a boolean
2045 /// indicating whether any of the contributing operations will actually
2046 /// operate on
2047 /// vector values after type legalization in the backend. If this latter value
2048 /// is
2049 /// false, then all operations will be scalarized (i.e. no vectorization has
2050 /// actually taken place).
2051 typedef std::pair<unsigned, bool> VectorizationCostTy;
2052
2053 /// Returns the expected execution cost. The unit of the cost does
2054 /// not matter because we use the 'cost' units to compare different
2055 /// vector widths. The cost that is returned is *not* normalized by
2056 /// the factor width.
2057 VectorizationCostTy expectedCost(unsigned VF);
2058
2059 /// Returns the execution time cost of an instruction for a given vector
2060 /// width. Vector width of one means scalar.
2061 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
2062
2063 /// The cost-computation logic from getInstructionCost which provides
2064 /// the vector type as an output parameter.
2065 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
2066
2067 /// Calculate vectorization cost of memory instruction \p I.
2068 unsigned getMemoryInstructionCost(Instruction *I, unsigned VF);
2069
2070 /// The cost computation for scalarized memory instruction.
2071 unsigned getMemInstScalarizationCost(Instruction *I, unsigned VF);
2072
2073 /// The cost computation for interleaving group of memory instructions.
2074 unsigned getInterleaveGroupCost(Instruction *I, unsigned VF);
2075
2076 /// The cost computation for Gather/Scatter instruction.
2077 unsigned getGatherScatterCost(Instruction *I, unsigned VF);
2078
2079 /// The cost computation for widening instruction \p I with consecutive
2080 /// memory access.
2081 unsigned getConsecutiveMemOpCost(Instruction *I, unsigned VF);
2082
2083 /// The cost calculation for Load instruction \p I with uniform pointer -
2084 /// scalar load + broadcast.
2085 unsigned getUniformMemOpCost(Instruction *I, unsigned VF);
2086
2087 /// Returns whether the instruction is a load or store and will be a emitted
2088 /// as a vector operation.
2089 bool isConsecutiveLoadOrStore(Instruction *I);
2090
2091 /// Create an analysis remark that explains why vectorization failed
2092 ///
2093 /// \p RemarkName is the identifier for the remark. \return the remark object
2094 /// that can be streamed to.
2095 OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
2096 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
2097 RemarkName, TheLoop);
2098 }
2099
2100 /// Map of scalar integer values to the smallest bitwidth they can be legally
2101 /// represented as. The vector equivalents of these values should be truncated
2102 /// to this type.
2103 MapVector<Instruction *, uint64_t> MinBWs;
2104
2105 /// A type representing the costs for instructions if they were to be
2106 /// scalarized rather than vectorized. The entries are Instruction-Cost
2107 /// pairs.
2108 typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
2109
2110 /// A map holding scalar costs for different vectorization factors. The
2111 /// presence of a cost for an instruction in the mapping indicates that the
2112 /// instruction will be scalarized when vectorizing with the associated
2113 /// vectorization factor. The entries are VF-ScalarCostTy pairs.
2114 DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
2115
2116 /// Holds the instructions known to be uniform after vectorization.
2117 /// The data is collected per VF.
2118 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Uniforms;
2119
2120 /// Holds the instructions known to be scalar after vectorization.
2121 /// The data is collected per VF.
2122 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Scalars;
2123
2124 /// Returns the expected difference in cost from scalarizing the expression
2125 /// feeding a predicated instruction \p PredInst. The instructions to
2126 /// scalarize and their scalar costs are collected in \p ScalarCosts. A
2127 /// non-negative return value implies the expression will be scalarized.
2128 /// Currently, only single-use chains are considered for scalarization.
2129 int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
2130 unsigned VF);
2131
2132 /// Collects the instructions to scalarize for each predicated instruction in
2133 /// the loop.
2134 void collectInstsToScalarize(unsigned VF);
2135
2136 /// Collect the instructions that are uniform after vectorization. An
2137 /// instruction is uniform if we represent it with a single scalar value in
2138 /// the vectorized loop corresponding to each vector iteration. Examples of
2139 /// uniform instructions include pointer operands of consecutive or
2140 /// interleaved memory accesses. Note that although uniformity implies an
2141 /// instruction will be scalar, the reverse is not true. In general, a
2142 /// scalarized instruction will be represented by VF scalar values in the
2143 /// vectorized loop, each corresponding to an iteration of the original
2144 /// scalar loop.
2145 void collectLoopUniforms(unsigned VF);
2146
2147 /// Collect the instructions that are scalar after vectorization. An
2148 /// instruction is scalar if it is known to be uniform or will be scalarized
2149 /// during vectorization. Non-uniform scalarized instructions will be
2150 /// represented by VF values in the vectorized loop, each corresponding to an
2151 /// iteration of the original scalar loop.
2152 void collectLoopScalars(unsigned VF);
2153
2154 /// Collect Uniform and Scalar values for the given \p VF.
2155 /// The sets depend on CM decision for Load/Store instructions
2156 /// that may be vectorized as interleave, gather-scatter or scalarized.
2157 void collectUniformsAndScalars(unsigned VF) {
2158 // Do the analysis once.
2159 if (VF == 1 || Uniforms.count(VF))
2160 return;
2161 setCostBasedWideningDecision(VF);
2162 collectLoopUniforms(VF);
2163 collectLoopScalars(VF);
2164 }
2165
2166 /// Keeps cost model vectorization decision and cost for instructions.
2167 /// Right now it is used for memory instructions only.
2168 typedef DenseMap<std::pair<Instruction *, unsigned>,
2169 std::pair<InstWidening, unsigned>>
2170 DecisionList;
2171
2172 DecisionList WideningDecisions;
2173
2174public:
2175 /// The loop that we evaluate.
2176 Loop *TheLoop;
2177 /// Predicated scalar evolution analysis.
2178 PredicatedScalarEvolution &PSE;
2179 /// Loop Info analysis.
2180 LoopInfo *LI;
2181 /// Vectorization legality.
2182 LoopVectorizationLegality *Legal;
2183 /// Vector target information.
2184 const TargetTransformInfo &TTI;
2185 /// Target Library Info.
2186 const TargetLibraryInfo *TLI;
2187 /// Demanded bits analysis.
2188 DemandedBits *DB;
2189 /// Assumption cache.
2190 AssumptionCache *AC;
2191 /// Interface to emit optimization remarks.
2192 OptimizationRemarkEmitter *ORE;
2193
2194 const Function *TheFunction;
2195 /// Loop Vectorize Hint.
2196 const LoopVectorizeHints *Hints;
2197 /// Values to ignore in the cost model.
2198 SmallPtrSet<const Value *, 16> ValuesToIgnore;
2199 /// Values to ignore in the cost model when VF > 1.
2200 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2201};
2202
2203/// \brief This holds vectorization requirements that must be verified late in
2204/// the process. The requirements are set by legalize and costmodel. Once
2205/// vectorization has been determined to be possible and profitable the
2206/// requirements can be verified by looking for metadata or compiler options.
2207/// For example, some loops require FP commutativity which is only allowed if
2208/// vectorization is explicitly specified or if the fast-math compiler option
2209/// has been provided.
2210/// Late evaluation of these requirements allows helpful diagnostics to be
2211/// composed that tells the user what need to be done to vectorize the loop. For
2212/// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2213/// evaluation should be used only when diagnostics can generated that can be
2214/// followed by a non-expert user.
2215class LoopVectorizationRequirements {
2216public:
2217 LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE)
2218 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr), ORE(ORE) {}
2219
2220 void addUnsafeAlgebraInst(Instruction *I) {
2221 // First unsafe algebra instruction.
2222 if (!UnsafeAlgebraInst)
2223 UnsafeAlgebraInst = I;
2224 }
2225
2226 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2227
2228 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2229 const char *PassName = Hints.vectorizeAnalysisPassName();
2230 bool Failed = false;
2231 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2232 ORE.emit(
2233 OptimizationRemarkAnalysisFPCommute(PassName, "CantReorderFPOps",
2234 UnsafeAlgebraInst->getDebugLoc(),
2235 UnsafeAlgebraInst->getParent())
2236 << "loop not vectorized: cannot prove it is safe to reorder "
2237 "floating-point operations");
2238 Failed = true;
2239 }
2240
2241 // Test if runtime memcheck thresholds are exceeded.
2242 bool PragmaThresholdReached =
2243 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2244 bool ThresholdReached =
2245 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2246 if ((ThresholdReached && !Hints.allowReordering()) ||
2247 PragmaThresholdReached) {
2248 ORE.emit(OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2249 L->getStartLoc(),
2250 L->getHeader())
2251 << "loop not vectorized: cannot prove it is safe to reorder "
2252 "memory operations");
2253 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)
;
2254 Failed = true;
2255 }
2256
2257 return Failed;
2258 }
2259
2260private:
2261 unsigned NumRuntimePointerChecks;
2262 Instruction *UnsafeAlgebraInst;
2263
2264 /// Interface to emit optimization remarks.
2265 OptimizationRemarkEmitter &ORE;
2266};
2267
2268static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
2269 if (L.empty()) {
2270 if (!hasCyclesInLoopBody(L))
2271 V.push_back(&L);
2272 return;
2273 }
2274 for (Loop *InnerL : L)
2275 addAcyclicInnerLoop(*InnerL, V);
2276}
2277
2278/// The LoopVectorize Pass.
2279struct LoopVectorize : public FunctionPass {
2280 /// Pass identification, replacement for typeid
2281 static char ID;
2282
2283 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2284 : FunctionPass(ID) {
2285 Impl.DisableUnrolling = NoUnrolling;
2286 Impl.AlwaysVectorize = AlwaysVectorize;
2287 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
2288 }
2289
2290 LoopVectorizePass Impl;
2291
2292 bool runOnFunction(Function &F) override {
2293 if (skipFunction(F))
2294 return false;
2295
2296 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2297 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2298 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2299 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2300 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2301 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2302 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2303 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2304 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2305 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2306 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2307 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2308
2309 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2310 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2311
2312 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2313 GetLAA, *ORE);
2314 }
2315
2316 void getAnalysisUsage(AnalysisUsage &AU) const override {
2317 AU.addRequired<AssumptionCacheTracker>();
2318 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2319 AU.addRequired<DominatorTreeWrapperPass>();
2320 AU.addRequired<LoopInfoWrapperPass>();
2321 AU.addRequired<ScalarEvolutionWrapperPass>();
2322 AU.addRequired<TargetTransformInfoWrapperPass>();
2323 AU.addRequired<AAResultsWrapperPass>();
2324 AU.addRequired<LoopAccessLegacyAnalysis>();
2325 AU.addRequired<DemandedBitsWrapperPass>();
2326 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2327 AU.addPreserved<LoopInfoWrapperPass>();
2328 AU.addPreserved<DominatorTreeWrapperPass>();
2329 AU.addPreserved<BasicAAWrapperPass>();
2330 AU.addPreserved<GlobalsAAWrapperPass>();
2331 }
2332};
2333
2334} // end anonymous namespace
2335
2336//===----------------------------------------------------------------------===//
2337// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2338// LoopVectorizationCostModel.
2339//===----------------------------------------------------------------------===//
2340
2341Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
2342 // We need to place the broadcast of invariant variables outside the loop.
2343 Instruction *Instr = dyn_cast<Instruction>(V);
2344 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2345 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2346
2347 // Place the code for broadcasting invariant variables in the new preheader.
2348 IRBuilder<>::InsertPointGuard Guard(Builder);
2349 if (Invariant)
2350 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2351
2352 // Broadcast the scalar into all locations in the vector.
2353 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2354
2355 return Shuf;
2356}
2357
2358void InnerLoopVectorizer::createVectorIntInductionPHI(
2359 const InductionDescriptor &II, Value *Step, Instruction *EntryVal) {
2360 Value *Start = II.getStartValue();
2361 assert(Step->getType()->isIntegerTy() &&((Step->getType()->isIntegerTy() && "Cannot widen an IV having a step with a non-integer type"
) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Cannot widen an IV having a step with a non-integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2362, __PRETTY_FUNCTION__))
2362 "Cannot widen an IV having a step with a non-integer type")((Step->getType()->isIntegerTy() && "Cannot widen an IV having a step with a non-integer type"
) ? static_cast<void> (0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Cannot widen an IV having a step with a non-integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2362, __PRETTY_FUNCTION__))
;
2363
2364 // Construct the initial value of the vector IV in the vector loop preheader
2365 auto CurrIP = Builder.saveIP();
2366 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2367 if (isa<TruncInst>(EntryVal)) {
2368 auto *TruncType = cast<IntegerType>(EntryVal->getType());
2369 Step = Builder.CreateTrunc(Step, TruncType);
2370 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2371 }
2372 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2373 Value *SteppedStart = getStepVector(SplatStart, 0, Step);
2374
2375 // Create a vector splat to use in the induction update.
2376 //
2377 // FIXME: If the step is non-constant, we create the vector splat with
2378 // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
2379 // handle a constant vector splat.
2380 auto *ConstVF = ConstantInt::getSigned(Step->getType(), VF);
2381 auto *Mul = Builder.CreateMul(Step, ConstVF);
2382 Value *SplatVF = isa<Constant>(Mul)
2383 ? ConstantVector::getSplat(VF, cast<Constant>(Mul))
2384 : Builder.CreateVectorSplat(VF, Mul);
2385 Builder.restoreIP(CurrIP);
2386
2387 // We may need to add the step a number of times, depending on the unroll
2388 // factor. The last of those goes into the PHI.
2389 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2390 &*LoopVectorBody->getFirstInsertionPt());
2391 Instruction *LastInduction = VecInd;
2392 VectorParts Entry(UF);
2393 for (unsigned Part = 0; Part < UF; ++Part) {
2394 Entry[Part] = LastInduction;
2395 LastInduction = cast<Instruction>(
2396 Builder.CreateAdd(LastInduction, SplatVF, "step.add"));
2397 }
2398 VectorLoopValueMap.initVector(EntryVal, Entry);
2399 if (isa<TruncInst>(EntryVal))
2400 addMetadata(Entry, EntryVal);
2401
2402 // Move the last step to the end of the latch block. This ensures consistent
2403 // placement of all induction updates.
2404 auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2405 auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2406 auto *ICmp = cast<Instruction>(Br->getCondition());
2407 LastInduction->moveBefore(ICmp);
2408 LastInduction->setName("vec.ind.next");
2409
2410 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2411 VecInd->addIncoming(LastInduction, LoopVectorLatch);
2412}
2413
2414bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
2415 return Cost->isScalarAfterVectorization(I, VF) ||
2416 Cost->isProfitableToScalarize(I, VF);
2417}
2418
2419bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
2420 if (shouldScalarizeInstruction(IV))
2421 return true;
2422 auto isScalarInst = [&](User *U) -> bool {
2423 auto *I = cast<Instruction>(U);
2424 return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2425 };
2426 return any_of(IV->users(), isScalarInst);
2427}
2428
2429void InnerLoopVectorizer::widenIntInduction(PHINode *IV, TruncInst *Trunc) {
2430
2431 auto II = Legal->getInductionVars()->find(IV);
2432 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2432, __PRETTY_FUNCTION__))
;
2433
2434 auto ID = II->second;
2435 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2435, __PRETTY_FUNCTION__))
;
2436
2437 // The scalar value to broadcast. This will be derived from the canonical
2438 // induction variable.
2439 Value *ScalarIV = nullptr;
2440
2441 // The value from the original loop to which we are mapping the new induction
2442 // variable.
2443 Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2444
2445 // True if we have vectorized the induction variable.
2446 auto VectorizedIV = false;
2447
2448 // Determine if we want a scalar version of the induction variable. This is
2449 // true if the induction variable itself is not widened, or if it has at
2450 // least one user in the loop that is not widened.
2451 auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2452
2453 // Generate code for the induction step. Note that induction steps are
2454 // required to be loop-invariant
2455 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2456, __PRETTY_FUNCTION__))
2456 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2456, __PRETTY_FUNCTION__))
;
2457 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2458 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2459 Value *Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2460 LoopVectorPreHeader->getTerminator());
2461
2462 // Try to create a new independent vector induction variable. If we can't
2463 // create the phi node, we will splat the scalar induction variable in each
2464 // loop iteration.
2465 if (VF > 1 && !shouldScalarizeInstruction(EntryVal)) {
2466 createVectorIntInductionPHI(ID, Step, EntryVal);
2467 VectorizedIV = true;
2468 }
2469
2470 // If we haven't yet vectorized the induction variable, or if we will create
2471 // a scalar one, we need to define the scalar induction variable and step
2472 // values. If we were given a truncation type, truncate the canonical
2473 // induction variable and step. Otherwise, derive these values from the
2474 // induction descriptor.
2475 if (!VectorizedIV || NeedsScalarIV) {
2476 if (Trunc) {
2477 auto *TruncType = cast<IntegerType>(Trunc->getType());
2478 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2479, __PRETTY_FUNCTION__))
2479 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2479, __PRETTY_FUNCTION__))
;
2480 ScalarIV = Builder.CreateCast(Instruction::Trunc, Induction, TruncType);
2481 Step = Builder.CreateTrunc(Step, TruncType);
2482 } else {
2483 ScalarIV = Induction;
2484 if (IV != OldInduction) {
2485 ScalarIV = Builder.CreateSExtOrTrunc(ScalarIV, IV->getType());
2486 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2487 ScalarIV->setName("offset.idx");
2488 }
2489 }
2490 }
2491
2492 // If we haven't yet vectorized the induction variable, splat the scalar
2493 // induction variable, and build the necessary step vectors.
2494 if (!VectorizedIV) {
2495 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2496 VectorParts Entry(UF);
2497 for (unsigned Part = 0; Part < UF; ++Part)
2498 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
2499 VectorLoopValueMap.initVector(EntryVal, Entry);
2500 if (Trunc)
2501 addMetadata(Entry, Trunc);
2502 }
2503
2504 // If an induction variable is only used for counting loop iterations or
2505 // calculating addresses, it doesn't need to be widened. Create scalar steps
2506 // that can be used by instructions we will later scalarize. Note that the
2507 // addition of the scalar steps will not increase the number of instructions
2508 // in the loop in the common case prior to InstCombine. We will be trading
2509 // one vector extract for each scalar step.
2510 if (NeedsScalarIV)
2511 buildScalarSteps(ScalarIV, Step, EntryVal);
2512}
2513
2514Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
2515 Instruction::BinaryOps BinOp) {
2516 // Create and check the types.
2517 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2517, __PRETTY_FUNCTION__))
;
2518 int VLen = Val->getType()->getVectorNumElements();
2519
2520 Type *STy = Val->getType()->getScalarType();
2521 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2522, __PRETTY_FUNCTION__))
2522 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2522, __PRETTY_FUNCTION__))
;
2523 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2523, __PRETTY_FUNCTION__))
;
2524
2525 SmallVector<Constant *, 8> Indices;
2526
2527 if (STy->isIntegerTy()) {
2528 // Create a vector of consecutive numbers from zero to VF.
2529 for (int i = 0; i < VLen; ++i)
2530 Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2531
2532 // Add the consecutive indices to the vector value.
2533 Constant *Cv = ConstantVector::get(Indices);
2534 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2534, __PRETTY_FUNCTION__))
;
2535 Step = Builder.CreateVectorSplat(VLen, Step);
2536 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2536, __PRETTY_FUNCTION__))
;
2537 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2538 // which can be found from the original scalar operations.
2539 Step = Builder.CreateMul(Cv, Step);
2540 return Builder.CreateAdd(Val, Step, "induction");
2541 }
2542
2543 // Floating point induction.
2544 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2545, __PRETTY_FUNCTION__))
2545 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2545, __PRETTY_FUNCTION__))
;
2546 // Create a vector of consecutive numbers from zero to VF.
2547 for (int i = 0; i < VLen; ++i)
2548 Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2549
2550 // Add the consecutive indices to the vector value.
2551 Constant *Cv = ConstantVector::get(Indices);
2552
2553 Step = Builder.CreateVectorSplat(VLen, Step);
2554
2555 // Floating point operations had to be 'fast' to enable the induction.
2556 FastMathFlags Flags;
2557 Flags.setUnsafeAlgebra();
2558
2559 Value *MulOp = Builder.CreateFMul(Cv, Step);
2560 if (isa<Instruction>(MulOp))
2561 // Have to check, MulOp may be a constant
2562 cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2563
2564 Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2565 if (isa<Instruction>(BOp))
2566 cast<Instruction>(BOp)->setFastMathFlags(Flags);
2567 return BOp;
2568}
2569
2570void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2571 Value *EntryVal) {
2572
2573 // We shouldn't have to build scalar steps if we aren't vectorizing.
2574 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2574, __PRETTY_FUNCTION__))
;
2575
2576 // Get the value type and ensure it and the step have the same integer type.
2577 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2578 assert(ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() &&((ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->
getType() && "Val and Step should have the same integer type"
) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() && \"Val and Step should have the same integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2579, __PRETTY_FUNCTION__))
2579 "Val and Step should have the same integer type")((ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->
getType() && "Val and Step should have the same integer type"
) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() && \"Val and Step should have the same integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2579, __PRETTY_FUNCTION__))
;
2580
2581 // Determine the number of scalars we need to generate for each unroll
2582 // iteration. If EntryVal is uniform, we only need to generate the first
2583 // lane. Otherwise, we generate all VF values.
2584 unsigned Lanes =
2585 Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) ? 1 : VF;
2586
2587 // Compute the scalar steps and save the results in VectorLoopValueMap.
2588 ScalarParts Entry(UF);
2589 for (unsigned Part = 0; Part < UF; ++Part) {
2590 Entry[Part].resize(VF);
2591 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2592 auto *StartIdx = ConstantInt::get(ScalarIVTy, VF * Part + Lane);
2593 auto *Mul = Builder.CreateMul(StartIdx, Step);
2594 auto *Add = Builder.CreateAdd(ScalarIV, Mul);
2595 Entry[Part][Lane] = Add;
2596 }
2597 }
2598 VectorLoopValueMap.initScalar(EntryVal, Entry);
2599}
2600
2601int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2602
2603 const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2604 ValueToValueMap();
2605
2606 int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2607 if (Stride == 1 || Stride == -1)
2608 return Stride;
2609 return 0;
2610}
2611
2612bool LoopVectorizationLegality::isUniform(Value *V) {
2613 return LAI->isUniform(V);
2614}
2615
2616const InnerLoopVectorizer::VectorParts &
2617InnerLoopVectorizer::getVectorValue(Value *V) {
2618 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2618, __PRETTY_FUNCTION__))
;
2619 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2619, __PRETTY_FUNCTION__))
;
2620 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2620, __PRETTY_FUNCTION__))
;
2621
2622 // If we have a stride that is replaced by one, do it here.
2623 if (Legal->hasStride(V))
2624 V = ConstantInt::get(V->getType(), 1);
2625
2626 // If we have this scalar in the map, return it.
2627 if (VectorLoopValueMap.hasVector(V))
2628 return VectorLoopValueMap.VectorMapStorage[V];
2629
2630 // If the value has not been vectorized, check if it has been scalarized
2631 // instead. If it has been scalarized, and we actually need the value in
2632 // vector form, we will construct the vector values on demand.
2633 if (VectorLoopValueMap.hasScalar(V)) {
2634
2635 // Initialize a new vector map entry.
2636 VectorParts Entry(UF);
2637
2638 // If we've scalarized a value, that value should be an instruction.
2639 auto *I = cast<Instruction>(V);
2640
2641 // If we aren't vectorizing, we can just copy the scalar map values over to
2642 // the vector map.
2643 if (VF == 1) {
2644 for (unsigned Part = 0; Part < UF; ++Part)
2645 Entry[Part] = getScalarValue(V, Part, 0);
2646 return VectorLoopValueMap.initVector(V, Entry);
2647 }
2648
2649 // Get the last scalar instruction we generated for V. If the value is
2650 // known to be uniform after vectorization, this corresponds to lane zero
2651 // of the last unroll iteration. Otherwise, the last instruction is the one
2652 // we created for the last vector lane of the last unroll iteration.
2653 unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) ? 0 : VF - 1;
2654 auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
2655
2656 // Set the insert point after the last scalarized instruction. This ensures
2657 // the insertelement sequence will directly follow the scalar definitions.
2658 auto OldIP = Builder.saveIP();
2659 auto NewIP = std::next(BasicBlock::iterator(LastInst));
2660 Builder.SetInsertPoint(&*NewIP);
2661
2662 // However, if we are vectorizing, we need to construct the vector values.
2663 // If the value is known to be uniform after vectorization, we can just
2664 // broadcast the scalar value corresponding to lane zero for each unroll
2665 // iteration. Otherwise, we construct the vector values using insertelement
2666 // instructions. Since the resulting vectors are stored in
2667 // VectorLoopValueMap, we will only generate the insertelements once.
2668 for (unsigned Part = 0; Part < UF; ++Part) {
2669 Value *VectorValue = nullptr;
2670 if (Cost->isUniformAfterVectorization(I, VF)) {
2671 VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
2672 } else {
2673 VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
2674 for (unsigned Lane = 0; Lane < VF; ++Lane)
2675 VectorValue = Builder.CreateInsertElement(
2676 VectorValue, getScalarValue(V, Part, Lane),
2677 Builder.getInt32(Lane));
2678 }
2679 Entry[Part] = VectorValue;
2680 }
2681 Builder.restoreIP(OldIP);
2682 return VectorLoopValueMap.initVector(V, Entry);
2683 }
2684
2685 // If this scalar is unknown, assume that it is a constant or that it is
2686 // loop invariant. Broadcast V and save the value for future uses.
2687 Value *B = getBroadcastInstrs(V);
2688 return VectorLoopValueMap.initVector(V, VectorParts(UF, B));
2689}
2690
2691Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
2692 unsigned Lane) {
2693
2694 // If the value is not an instruction contained in the loop, it should
2695 // already be scalar.
2696 if (OrigLoop->isLoopInvariant(V))
2697 return V;
2698
2699 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2701, __PRETTY_FUNCTION__))
2700 !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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2701, __PRETTY_FUNCTION__))
2701 : 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2701, __PRETTY_FUNCTION__))
;
2702
2703 // If the value from the original loop has not been vectorized, it is
2704 // represented by UF x VF scalar values in the new loop. Return the requested
2705 // scalar value.
2706 if (VectorLoopValueMap.hasScalar(V))
2707 return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
2708
2709 // If the value has not been scalarized, get its entry in VectorLoopValueMap
2710 // for the given unroll part. If this entry is not a vector type (i.e., the
2711 // vectorization factor is one), there is no need to generate an
2712 // extractelement instruction.
2713 auto *U = getVectorValue(V)[Part];
2714 if (!U->getType()->isVectorTy()) {
2715 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2715, __PRETTY_FUNCTION__))
;
2716 return U;
2717 }
2718
2719 // Otherwise, the value from the original loop has been vectorized and is
2720 // represented by UF vector values. Extract and return the requested scalar
2721 // value from the appropriate vector lane.
2722 return Builder.CreateExtractElement(U, Builder.getInt32(Lane));
2723}
2724
2725Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2726 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2726, __PRETTY_FUNCTION__))
;
2727 SmallVector<Constant *, 8> ShuffleMask;
2728 for (unsigned i = 0; i < VF; ++i)
2729 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2730
2731 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2732 ConstantVector::get(ShuffleMask),
2733 "reverse");
2734}
2735
2736// Try to vectorize the interleave group that \p Instr belongs to.
2737//
2738// E.g. Translate following interleaved load group (factor = 3):
2739// for (i = 0; i < N; i+=3) {
2740// R = Pic[i]; // Member of index 0
2741// G = Pic[i+1]; // Member of index 1
2742// B = Pic[i+2]; // Member of index 2
2743// ... // do something to R, G, B
2744// }
2745// To:
2746// %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2747// %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2748// %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2749// %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2750//
2751// Or translate following interleaved store group (factor = 3):
2752// for (i = 0; i < N; i+=3) {
2753// ... do something to R, G, B
2754// Pic[i] = R; // Member of index 0
2755// Pic[i+1] = G; // Member of index 1
2756// Pic[i+2] = B; // Member of index 2
2757// }
2758// To:
2759// %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2760// %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2761// %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2762// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2763// store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2764void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2765 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2766 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2766, __PRETTY_FUNCTION__))
;
2767
2768 // Skip if current instruction is not the insert position.
2769 if (Instr != Group->getInsertPos())
2770 return;
2771
2772 Value *Ptr = getPointerOperand(Instr);
2773
2774 // Prepare for the vector type of the interleaved load/store.
2775 Type *ScalarTy = getMemInstValueType(Instr);
2776 unsigned InterleaveFactor = Group->getFactor();
2777 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2778 Type *PtrTy = VecTy->getPointerTo(getMemInstAddressSpace(Instr));
2779
2780 // Prepare for the new pointers.
2781 setDebugLocFromInst(Builder, Ptr);
2782 SmallVector<Value *, 2> NewPtrs;
2783 unsigned Index = Group->getIndex(Instr);
2784
2785 // If the group is reverse, adjust the index to refer to the last vector lane
2786 // instead of the first. We adjust the index from the first vector lane,
2787 // rather than directly getting the pointer for lane VF - 1, because the
2788 // pointer operand of the interleaved access is supposed to be uniform. For
2789 // uniform instructions, we're only required to generate a value for the
2790 // first vector lane in each unroll iteration.
2791 if (Group->isReverse())
2792 Index += (VF - 1) * Group->getFactor();
2793
2794 for (unsigned Part = 0; Part < UF; Part++) {
2795 Value *NewPtr = getScalarValue(Ptr, Part, 0);
2796
2797 // Notice current instruction could be any index. Need to adjust the address
2798 // to the member of index 0.
2799 //
2800 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2801 // b = A[i]; // Member of index 0
2802 // Current pointer is pointed to A[i+1], adjust it to A[i].
2803 //
2804 // E.g. A[i+1] = a; // Member of index 1
2805 // A[i] = b; // Member of index 0
2806 // A[i+2] = c; // Member of index 2 (Current instruction)
2807 // Current pointer is pointed to A[i+2], adjust it to A[i].
2808 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2809
2810 // Cast to the vector pointer type.
2811 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2812 }
2813
2814 setDebugLocFromInst(Builder, Instr);
2815 Value *UndefVec = UndefValue::get(VecTy);
2816
2817 // Vectorize the interleaved load group.
2818 if (isa<LoadInst>(Instr)) {
2819
2820 // For each unroll part, create a wide load for the group.
2821 SmallVector<Value *, 2> NewLoads;
2822 for (unsigned Part = 0; Part < UF; Part++) {
2823 auto *NewLoad = Builder.CreateAlignedLoad(
2824 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2825 addMetadata(NewLoad, Instr);
2826 NewLoads.push_back(NewLoad);
2827 }
2828
2829 // For each member in the group, shuffle out the appropriate data from the
2830 // wide loads.
2831 for (unsigned I = 0; I < InterleaveFactor; ++I) {
2832 Instruction *Member = Group->getMember(I);
2833
2834 // Skip the gaps in the group.
2835 if (!Member)
2836 continue;
2837
2838 VectorParts Entry(UF);
2839 Constant *StrideMask = createStrideMask(Builder, I, InterleaveFactor, VF);
2840 for (unsigned Part = 0; Part < UF; Part++) {
2841 Value *StridedVec = Builder.CreateShuffleVector(
2842 NewLoads[Part], UndefVec, StrideMask, "strided.vec");
2843
2844 // If this member has different type, cast the result type.
2845 if (Member->getType() != ScalarTy) {
2846 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2847 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2848 }
2849
2850 Entry[Part] =
2851 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2852 }
2853 VectorLoopValueMap.initVector(Member, Entry);
2854 }
2855 return;
2856 }
2857
2858 // The sub vector type for current instruction.
2859 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2860
2861 // Vectorize the interleaved store group.
2862 for (unsigned Part = 0; Part < UF; Part++) {
2863 // Collect the stored vector from each member.
2864 SmallVector<Value *, 4> StoredVecs;
2865 for (unsigned i = 0; i < InterleaveFactor; i++) {
2866 // Interleaved store group doesn't allow a gap, so each index has a member
2867 Instruction *Member = Group->getMember(i);
2868 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2868, __PRETTY_FUNCTION__))
;
2869
2870 Value *StoredVec =
2871 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2872 if (Group->isReverse())
2873 StoredVec = reverseVector(StoredVec);
2874
2875 // If this member has different type, cast it to an unified type.
2876 if (StoredVec->getType() != SubVT)
2877 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2878
2879 StoredVecs.push_back(StoredVec);
2880 }
2881
2882 // Concatenate all vectors into a wide vector.
2883 Value *WideVec = concatenateVectors(Builder, StoredVecs);
2884
2885 // Interleave the elements in the wide vector.
2886 Constant *IMask = createInterleaveMask(Builder, VF, InterleaveFactor);
2887 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2888 "interleaved.vec");
2889
2890 Instruction *NewStoreInstr =
2891 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2892 addMetadata(NewStoreInstr, Instr);
2893 }
2894}
2895
2896void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2897 // Attempt to issue a wide load.
2898 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2899 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2900
2901 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2901, __PRETTY_FUNCTION__))
;
2902
2903 LoopVectorizationCostModel::InstWidening Decision =
2904 Cost->getWideningDecision(Instr, VF);
2905 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2906, __PRETTY_FUNCTION__))
2906 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2906, __PRETTY_FUNCTION__))
;
2907 if (Decision == LoopVectorizationCostModel::CM_Interleave)
2908 return vectorizeInterleaveGroup(Instr);
2909
2910 Type *ScalarDataTy = getMemInstValueType(Instr);
2911 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2912 Value *Ptr = getPointerOperand(Instr);
2913 unsigned Alignment = getMemInstAlignment(Instr);
2914 // An alignment of 0 means target abi alignment. We need to use the scalar's
2915 // target abi alignment in such a case.
2916 const DataLayout &DL = Instr->getModule()->getDataLayout();
2917 if (!Alignment)
2918 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2919 unsigned AddressSpace = getMemInstAddressSpace(Instr);
2920
2921 // Scalarize the memory instruction if necessary.
2922 if (Decision == LoopVectorizationCostModel::CM_Scalarize)
2923 return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
2924
2925 // Determine if the pointer operand of the access is either consecutive or
2926 // reverse consecutive.
2927 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2928 bool Reverse = ConsecutiveStride < 0;
2929 bool CreateGatherScatter =
2930 (Decision == LoopVectorizationCostModel::CM_GatherScatter);
2931
2932 VectorParts VectorGep;
2933
2934 // Handle consecutive loads/stores.
2935 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2936 if (ConsecutiveStride) {
2937 if (Gep) {
2938 unsigned NumOperands = Gep->getNumOperands();
2939#ifndef NDEBUG
2940 // The original GEP that identified as a consecutive memory access
2941 // should have only one loop-variant operand.
2942 unsigned NumOfLoopVariantOps = 0;
2943 for (unsigned i = 0; i < NumOperands; ++i)
2944 if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)),
2945 OrigLoop))
2946 NumOfLoopVariantOps++;
2947 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2948, __PRETTY_FUNCTION__))
2948 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2948, __PRETTY_FUNCTION__))
;
2949#endif
2950 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2951 Gep2->setName("gep.indvar");
2952
2953 // A new GEP is created for a 0-lane value of the first unroll iteration.
2954 // The GEPs for the rest of the unroll iterations are computed below as an
2955 // offset from this GEP.
2956 for (unsigned i = 0; i < NumOperands; ++i)
2957 // We can apply getScalarValue() for all GEP indices. It returns an
2958 // original value for loop-invariant operand and 0-lane for consecutive
2959 // operand.
2960 Gep2->setOperand(i, getScalarValue(Gep->getOperand(i),
2961 0, /* First unroll iteration */
2962 0 /* 0-lane of the vector */ ));
2963 setDebugLocFromInst(Builder, Gep);
2964 Ptr = Builder.Insert(Gep2);
2965
2966 } else { // No GEP
2967 setDebugLocFromInst(Builder, Ptr);
2968 Ptr = getScalarValue(Ptr, 0, 0);
2969 }
2970 } else {
2971 // At this point we should vector version of GEP for Gather or Scatter
2972 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2972, __PRETTY_FUNCTION__))
;
2973 if (Gep) {
2974 // Vectorizing GEP, across UF parts. We want to get a vector value for base
2975 // and each index that's defined inside the loop, even if it is
2976 // loop-invariant but wasn't hoisted out. Otherwise we want to keep them
2977 // scalar.
2978 SmallVector<VectorParts, 4> OpsV;
2979 for (Value *Op : Gep->operands()) {
2980 Instruction *SrcInst = dyn_cast<Instruction>(Op);
2981 if (SrcInst && OrigLoop->contains(SrcInst))
2982 OpsV.push_back(getVectorValue(Op));
2983 else
2984 OpsV.push_back(VectorParts(UF, Op));
2985 }
2986 for (unsigned Part = 0; Part < UF; ++Part) {
2987 SmallVector<Value *, 4> Ops;
2988 Value *GEPBasePtr = OpsV[0][Part];
2989 for (unsigned i = 1; i < Gep->getNumOperands(); i++)
2990 Ops.push_back(OpsV[i][Part]);
2991 Value *NewGep = Builder.CreateGEP(GEPBasePtr, Ops, "VectorGep");
2992 cast<GetElementPtrInst>(NewGep)->setIsInBounds(Gep->isInBounds());
2993 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2993, __PRETTY_FUNCTION__))
;
2994
2995 NewGep =
2996 Builder.CreateBitCast(NewGep, VectorType::get(Ptr->getType(), VF));
2997 VectorGep.push_back(NewGep);
2998 }
2999 } else
3000 VectorGep = getVectorValue(Ptr);
3001 }
3002
3003 VectorParts Mask = createBlockInMask(Instr->getParent());
3004 // Handle Stores:
3005 if (SI) {
3006 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3007, __PRETTY_FUNCTION__))
3007 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3007, __PRETTY_FUNCTION__))
;
3008 setDebugLocFromInst(Builder, SI);
3009 // We don't want to update the value in the map as it might be used in
3010 // another expression. So don't use a reference type for "StoredVal".
3011 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
3012
3013 for (unsigned Part = 0; Part < UF; ++Part) {
3014 Instruction *NewSI = nullptr;
3015 if (CreateGatherScatter) {
3016 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
3017 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
3018 Alignment, MaskPart);
3019 } else {
3020 // Calculate the pointer for the specific unroll-part.
3021 Value *PartPtr =
3022 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3023
3024 if (Reverse) {
3025 // If we store to reverse consecutive memory locations, then we need
3026 // to reverse the order of elements in the stored value.
3027 StoredVal[Part] = reverseVector(StoredVal[Part]);
3028 // If the address is consecutive but reversed, then the
3029 // wide store needs to start at the last vector element.
3030 PartPtr =
3031 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3032 PartPtr =
3033 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3034 Mask[Part] = reverseVector(Mask[Part]);
3035 }
3036
3037 Value *VecPtr =
3038 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3039
3040 if (Legal->isMaskRequired(SI))
3041 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
3042 Mask[Part]);
3043 else
3044 NewSI =
3045 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
3046 }
3047 addMetadata(NewSI, SI);
3048 }
3049 return;
3050 }
3051
3052 // Handle loads.
3053 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3053, __PRETTY_FUNCTION__))
;
3054 setDebugLocFromInst(Builder, LI);
3055 VectorParts Entry(UF);
3056 for (unsigned Part = 0; Part < UF; ++Part) {
3057 Instruction *NewLI;
3058 if (CreateGatherScatter) {
3059 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
3060 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
3061 0, "wide.masked.gather");
3062 Entry[Part] = NewLI;
3063 } else {
3064 // Calculate the pointer for the specific unroll-part.
3065 Value *PartPtr =
3066 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3067
3068 if (Reverse) {
3069 // If the address is consecutive but reversed, then the
3070 // wide load needs to start at the last vector element.
3071 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3072 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3073 Mask[Part] = reverseVector(Mask[Part]);
3074 }
3075
3076 Value *VecPtr =
3077 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3078 if (Legal->isMaskRequired(LI))
3079 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
3080 UndefValue::get(DataTy),
3081 "wide.masked.load");
3082 else
3083 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
3084 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
3085 }
3086 addMetadata(NewLI, LI);
3087 }
3088 VectorLoopValueMap.initVector(Instr, Entry);
3089}
3090
3091void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
3092 bool IfPredicateInstr) {
3093 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3093, __PRETTY_FUNCTION__))
;
3094 DEBUG(dbgs() << "LV: Scalarizing"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
3095 << (IfPredicateInstr ? " and predicating:" : ":") << *Instrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
3096 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
;
3097 // Holds vector parameters or scalars, in case of uniform vals.
3098 SmallVector<VectorParts, 4> Params;
3099
3100 setDebugLocFromInst(Builder, Instr);
3101
3102 // Does this instruction return a value ?
3103 bool IsVoidRetTy = Instr->getType()->isVoidTy();
3104
3105 // Initialize a new scalar map entry.
3106 ScalarParts Entry(UF);
3107
3108 VectorParts Cond;
3109 if (IfPredicateInstr)
3110 Cond = createBlockInMask(Instr->getParent());
3111
3112 // Determine the number of scalars we need to generate for each unroll
3113 // iteration. If the instruction is uniform, we only need to generate the
3114 // first lane. Otherwise, we generate all VF values.
3115 unsigned Lanes = Cost->isUniformAfterVectorization(Instr, VF) ? 1 : VF;
3116
3117 // For each vector unroll 'part':
3118 for (unsigned Part = 0; Part < UF; ++Part) {
3119 Entry[Part].resize(VF);
3120 // For each scalar that we create:
3121 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
3122
3123 // Start if-block.
3124 Value *Cmp = nullptr;
3125 if (IfPredicateInstr) {
3126 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Lane));
3127 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
3128 ConstantInt::get(Cmp->getType(), 1));
3129 }
3130
3131 Instruction *Cloned = Instr->clone();
3132 if (!IsVoidRetTy)
3133 Cloned->setName(Instr->getName() + ".cloned");
3134
3135 // Replace the operands of the cloned instructions with their scalar
3136 // equivalents in the new loop.
3137 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3138 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, Lane);
3139 Cloned->setOperand(op, NewOp);
3140 }
3141 addNewMetadata(Cloned, Instr);
3142
3143 // Place the cloned scalar in the new loop.
3144 Builder.Insert(Cloned);
3145
3146 // Add the cloned scalar to the scalar map entry.
3147 Entry[Part][Lane] = Cloned;
3148
3149 // If we just cloned a new assumption, add it the assumption cache.
3150 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3151 if (II->getIntrinsicID() == Intrinsic::assume)
3152 AC->registerAssumption(II);
3153
3154 // End if-block.
3155 if (IfPredicateInstr)
3156 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
3157 }
3158 }
3159 VectorLoopValueMap.initScalar(Instr, Entry);
3160}
3161
3162PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
3163 Value *End, Value *Step,
3164 Instruction *DL) {
3165 BasicBlock *Header = L->getHeader();
3166 BasicBlock *Latch = L->getLoopLatch();
3167 // As we're just creating this loop, it's possible no latch exists
3168 // yet. If so, use the header as this will be a single block loop.
3169 if (!Latch)
3170 Latch = Header;
3171
3172 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
3173 Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction);
3174 setDebugLocFromInst(Builder, OldInst);
3175 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3176
3177 Builder.SetInsertPoint(Latch->getTerminator());
3178 setDebugLocFromInst(Builder, OldInst);
3179
3180 // Create i+1 and fill the PHINode.
3181 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3182 Induction->addIncoming(Start, L->getLoopPreheader());
3183 Induction->addIncoming(Next, Latch);
3184 // Create the compare.
3185 Value *ICmp = Builder.CreateICmpEQ(Next, End);
3186 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3187
3188 // Now we have two terminators. Remove the old one from the block.
3189 Latch->getTerminator()->eraseFromParent();
3190
3191 return Induction;
3192}
3193
3194Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
3195 if (TripCount)
3196 return TripCount;
3197
3198 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3199 // Find the loop boundaries.
3200 ScalarEvolution *SE = PSE.getSE();
3201 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3202 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3203, __PRETTY_FUNCTION__))
3203 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3203, __PRETTY_FUNCTION__))
;
3204
3205 Type *IdxTy = Legal->getWidestInductionType();
3206
3207 // The exit count might have the type of i64 while the phi is i32. This can
3208 // happen if we have an induction variable that is sign extended before the
3209 // compare. The only way that we get a backedge taken count is that the
3210 // induction variable was signed and as such will not overflow. In such a case
3211 // truncation is legal.
3212 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3213 IdxTy->getPrimitiveSizeInBits())
3214 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3215 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3216
3217 // Get the total trip count from the count by adding 1.
3218 const SCEV *ExitCount = SE->getAddExpr(
3219 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3220
3221 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3222
3223 // Expand the trip count and place the new instructions in the preheader.
3224 // Notice that the pre-header does not change, only the loop body.
3225 SCEVExpander Exp(*SE, DL, "induction");
3226
3227 // Count holds the overall loop count (N).
3228 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3229 L->getLoopPreheader()->getTerminator());
3230
3231 if (TripCount->getType()->isPointerTy())
3232 TripCount =
3233 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3234 L->getLoopPreheader()->getTerminator());
3235
3236 return TripCount;
3237}
3238
3239Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
3240 if (VectorTripCount)
3241 return VectorTripCount;
3242
3243 Value *TC = getOrCreateTripCount(L);
3244 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3245
3246 // Now we need to generate the expression for the part of the loop that the
3247 // vectorized body will execute. This is equal to N - (N % Step) if scalar
3248 // iterations are not required for correctness, or N - Step, otherwise. Step
3249 // is equal to the vectorization factor (number of SIMD elements) times the
3250 // unroll factor (number of SIMD instructions).
3251 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3252 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3253
3254 // If there is a non-reversed interleaved group that may speculatively access
3255 // memory out-of-bounds, we need to ensure that there will be at least one
3256 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3257 // the trip count, we set the remainder to be equal to the step. If the step
3258 // does not evenly divide the trip count, no adjustment is necessary since
3259 // there will already be scalar iterations. Note that the minimum iterations
3260 // check ensures that N >= Step.
3261 if (VF > 1 && Legal->requiresScalarEpilogue()) {
3262 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3263 R = Builder.CreateSelect(IsZero, Step, R);
3264 }
3265
3266 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3267
3268 return VectorTripCount;
3269}
3270
3271void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
3272 BasicBlock *Bypass) {
3273 Value *Count = getOrCreateTripCount(L);
3274 BasicBlock *BB = L->getLoopPreheader();
3275 IRBuilder<> Builder(BB->getTerminator());
3276
3277 // Generate code to check that the loop's trip count that we computed by
3278 // adding one to the backedge-taken count will not overflow.
3279 Value *CheckMinIters = Builder.CreateICmpULT(
3280 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3281
3282 BasicBlock *NewBB =
3283 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
3284 // Update dominator tree immediately if the generated block is a
3285 // LoopBypassBlock because SCEV expansions to generate loop bypass
3286 // checks may query it before the current function is finished.
3287 DT->addNewBlock(NewBB, BB);
3288 if (L->getParentLoop())
3289 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3290 ReplaceInstWithInst(BB->getTerminator(),
3291 BranchInst::Create(Bypass, NewBB, CheckMinIters));
3292 LoopBypassBlocks.push_back(BB);
3293}
3294
3295void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
3296 BasicBlock *Bypass) {
3297 Value *TC = getOrCreateVectorTripCount(L);
3298 BasicBlock *BB = L->getLoopPreheader();
3299 IRBuilder<> Builder(BB->getTerminator());
3300
3301 // Now, compare the new count to zero. If it is zero skip the vector loop and
3302 // jump to the scalar loop.
3303 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
3304 "cmp.zero");
3305
3306 // Generate code to check that the loop's trip count that we computed by
3307 // adding one to the backedge-taken count will not overflow.
3308 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3309 // Update dominator tree immediately if the generated block is a
3310 // LoopBypassBlock because SCEV expansions to generate loop bypass
3311 // checks may query it before the current function is finished.
3312 DT->addNewBlock(NewBB, BB);
3313 if (L->getParentLoop())
3314 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3315 ReplaceInstWithInst(BB->getTerminator(),
3316 BranchInst::Create(Bypass, NewBB, Cmp));
3317 LoopBypassBlocks.push_back(BB);
3318}
3319
3320void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
3321 BasicBlock *BB = L->getLoopPreheader();
3322
3323 // Generate the code to check that the SCEV assumptions that we made.
3324 // We want the new basic block to start at the first instruction in a
3325 // sequence of instructions that form a check.
3326 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3327 "scev.check");
3328 Value *SCEVCheck =
3329 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3330
3331 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3332 if (C->isZero())
3333 return;
3334
3335 // Create a new block containing the stride check.
3336 BB->setName("vector.scevcheck");
3337 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3338 // Update dominator tree immediately if the generated block is a
3339 // LoopBypassBlock because SCEV expansions to generate loop bypass
3340 // checks may query it before the current function is finished.
3341 DT->addNewBlock(NewBB, BB);
3342 if (L->getParentLoop())
3343 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3344 ReplaceInstWithInst(BB->getTerminator(),
3345 BranchInst::Create(Bypass, NewBB, SCEVCheck));
3346 LoopBypassBlocks.push_back(BB);
3347 AddedSafetyChecks = true;
3348}
3349
3350void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
3351 BasicBlock *BB = L->getLoopPreheader();
3352
3353 // Generate the code that checks in runtime if arrays overlap. We put the
3354 // checks into a separate block to make the more common case of few elements
3355 // faster.
3356 Instruction *FirstCheckInst;
3357 Instruction *MemRuntimeCheck;
3358 std::tie(FirstCheckInst, MemRuntimeCheck) =
3359 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3360 if (!MemRuntimeCheck)
3361 return;
3362
3363 // Create a new block containing the memory check.
3364 BB->setName("vector.memcheck");
3365 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3366 // Update dominator tree immediately if the generated block is a
3367 // LoopBypassBlock because SCEV expansions to generate loop bypass
3368 // checks may query it before the current function is finished.
3369 DT->addNewBlock(NewBB, BB);
3370 if (L->getParentLoop())
3371 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3372 ReplaceInstWithInst(BB->getTerminator(),
3373 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3374 LoopBypassBlocks.push_back(BB);
3375 AddedSafetyChecks = true;
3376
3377 // We currently don't use LoopVersioning for the actual loop cloning but we
3378 // still use it to add the noalias metadata.
3379 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3380 PSE.getSE());
3381 LVer->prepareNoAliasMetadata();
3382}
3383
3384void InnerLoopVectorizer::createEmptyLoop() {
3385 /*
3386 In this function we generate a new loop. The new loop will contain
3387 the vectorized instructions while the old loop will continue to run the
3388 scalar remainder.
3389
3390 [ ] <-- loop iteration number check.
3391 / |
3392 / v
3393 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3394 | / |
3395 | / v
3396 || [ ] <-- vector pre header.
3397 |/ |
3398 | v
3399 | [ ] \
3400 | [ ]_| <-- vector loop.
3401 | |
3402 | v
3403 | -[ ] <--- middle-block.
3404 | / |
3405 | / v
3406 -|- >[ ] <--- new preheader.
3407 | |
3408 | v
3409 | [ ] \
3410 | [ ]_| <-- old scalar loop to handle remainder.
3411 \ |
3412 \ v
3413 >[ ] <-- exit block.
3414 ...
3415 */
3416
3417 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3418 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3419 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3420 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3420, __PRETTY_FUNCTION__))
;
3421 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3421, __PRETTY_FUNCTION__))
;
3422
3423 // Some loops have a single integer induction variable, while other loops
3424 // don't. One example is c++ iterators that often have multiple pointer
3425 // induction variables. In the code below we also support a case where we
3426 // don't have a single induction variable.
3427 //
3428 // We try to obtain an induction variable from the original loop as hard
3429 // as possible. However if we don't find one that:
3430 // - is an integer
3431 // - counts from zero, stepping by one
3432 // - is the size of the widest induction variable type
3433 // then we create a new one.
3434 OldInduction = Legal->getPrimaryInduction();
3435 Type *IdxTy = Legal->getWidestInductionType();
3436
3437 // Split the single block loop into the two loop structure described above.
3438 BasicBlock *VecBody =
3439 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3440 BasicBlock *MiddleBlock =
3441 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3442 BasicBlock *ScalarPH =
3443 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3444
3445 // Create and register the new vector loop.
3446 Loop *Lp = new Loop();
3447 Loop *ParentLoop = OrigLoop->getParentLoop();
3448
3449 // Insert the new loop into the loop nest and register the new basic blocks
3450 // before calling any utilities such as SCEV that require valid LoopInfo.
3451 if (ParentLoop) {
3452 ParentLoop->addChildLoop(Lp);
3453 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3454 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3455 } else {
3456 LI->addTopLevelLoop(Lp);
3457 }
3458 Lp->addBasicBlockToLoop(VecBody, *LI);
3459
3460 // Find the loop boundaries.
3461 Value *Count = getOrCreateTripCount(Lp);
3462
3463 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3464
3465 // We need to test whether the backedge-taken count is uint##_max. Adding one
3466 // to it will cause overflow and an incorrect loop trip count in the vector
3467 // body. In case of overflow we want to directly jump to the scalar remainder
3468 // loop.
3469 emitMinimumIterationCountCheck(Lp, ScalarPH);
3470 // Now, compare the new count to zero. If it is zero skip the vector loop and
3471 // jump to the scalar loop.
3472 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3473 // Generate the code to check any assumptions that we've made for SCEV
3474 // expressions.
3475 emitSCEVChecks(Lp, ScalarPH);
3476
3477 // Generate the code that checks in runtime if arrays overlap. We put the
3478 // checks into a separate block to make the more common case of few elements
3479 // faster.
3480 emitMemRuntimeChecks(Lp, ScalarPH);
3481
3482 // Generate the induction variable.
3483 // The loop step is equal to the vectorization factor (num of SIMD elements)
3484 // times the unroll factor (num of SIMD instructions).
3485 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3486 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3487 Induction =
3488 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3489 getDebugLocFromInstOrOperands(OldInduction));
3490
3491 // We are going to resume the execution of the scalar loop.
3492 // Go over all of the induction variables that we found and fix the
3493 // PHIs that are left in the scalar version of the loop.
3494 // The starting values of PHI nodes depend on the counter of the last
3495 // iteration in the vectorized loop.
3496 // If we come from a bypass edge then we need to start from the original
3497 // start value.
3498
3499 // This variable saves the new starting index for the scalar loop. It is used
3500 // to test if there are any tail iterations left once the vector loop has
3501 // completed.
3502 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3503 for (auto &InductionEntry : *List) {
3504 PHINode *OrigPhi = InductionEntry.first;
3505 InductionDescriptor II = InductionEntry.second;
3506
3507 // Create phi nodes to merge from the backedge-taken check block.
3508 PHINode *BCResumeVal = PHINode::Create(
3509 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3510 Value *&EndValue = IVEndValues[OrigPhi];
3511 if (OrigPhi == OldInduction) {
3512 // We know what the end value is.
3513 EndValue = CountRoundDown;
3514 } else {
3515 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3516 Type *StepType = II.getStep()->getType();
3517 Instruction::CastOps CastOp =
3518 CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3519 Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3520 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3521 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3522 EndValue->setName("ind.end");
3523 }
3524
3525 // The new PHI merges the original incoming value, in case of a bypass,
3526 // or the value at the end of the vectorized loop.
3527 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3528
3529 // Fix the scalar body counter (PHI node).
3530 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3531
3532 // The old induction's phi node in the scalar body needs the truncated
3533 // value.
3534 for (BasicBlock *BB : LoopBypassBlocks)
3535 BCResumeVal->addIncoming(II.getStartValue(), BB);
3536 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3537 }
3538
3539 // Add a check in the middle block to see if we have completed
3540 // all of the iterations in the first vector loop.
3541 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3542 Value *CmpN =
3543 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3544 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3545 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3546 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3547
3548 // Get ready to start creating new instructions into the vectorized body.
3549 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3550
3551 // Save the state.
3552 LoopVectorPreHeader = Lp->getLoopPreheader();
3553 LoopScalarPreHeader = ScalarPH;
3554 LoopMiddleBlock = MiddleBlock;
3555 LoopExitBlock = ExitBlock;
3556 LoopVectorBody = VecBody;
3557 LoopScalarBody = OldBasicBlock;
3558
3559 // Keep all loop hints from the original loop on the vector loop (we'll
3560 // replace the vectorizer-specific hints below).
3561 if (MDNode *LID = OrigLoop->getLoopID())
3562 Lp->setLoopID(LID);
3563
3564 LoopVectorizeHints Hints(Lp, true, *ORE);
3565 Hints.setAlreadyVectorized();
3566}
3567
3568// Fix up external users of the induction variable. At this point, we are
3569// in LCSSA form, with all external PHIs that use the IV having one input value,
3570// coming from the remainder loop. We need those PHIs to also have a correct
3571// value for the IV when arriving directly from the middle block.
3572void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3573 const InductionDescriptor &II,
3574 Value *CountRoundDown, Value *EndValue,
3575 BasicBlock *MiddleBlock) {
3576 // There are two kinds of external IV usages - those that use the value
3577 // computed in the last iteration (the PHI) and those that use the penultimate
3578 // value (the value that feeds into the phi from the loop latch).
3579 // We allow both, but they, obviously, have different values.
3580
3581 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3581, __PRETTY_FUNCTION__))
;
3582
3583 DenseMap<Value *, Value *> MissingVals;
3584
3585 // An external user of the last iteration's value should see the value that
3586 // the remainder loop uses to initialize its own IV.
3587 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3588 for (User *U : PostInc->users()) {
3589 Instruction *UI = cast<Instruction>(U);
3590 if (!OrigLoop->contains(UI)) {
3591 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3591, __PRETTY_FUNCTION__))
;
3592 MissingVals[UI] = EndValue;
3593 }
3594 }
3595
3596 // An external user of the penultimate value need to see EndValue - Step.
3597 // The simplest way to get this is to recompute it from the constituent SCEVs,
3598 // that is Start + (Step * (CRD - 1)).
3599 for (User *U : OrigPhi->users()) {
3600 auto *UI = cast<Instruction>(U);
3601 if (!OrigLoop->contains(UI)) {
3602 const DataLayout &DL =
3603 OrigLoop->getHeader()->getModule()->getDataLayout();
3604 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3604, __PRETTY_FUNCTION__))
;
3605
3606 IRBuilder<> B(MiddleBlock->getTerminator());
3607 Value *CountMinusOne = B.CreateSub(
3608 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3609 Value *CMO = B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType(),
3610 "cast.cmo");
3611 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3612 Escape->setName("ind.escape");
3613 MissingVals[UI] = Escape;
3614 }
3615 }
3616
3617 for (auto &I : MissingVals) {
3618 PHINode *PHI = cast<PHINode>(I.first);
3619 // One corner case we have to handle is two IVs "chasing" each-other,
3620 // that is %IV2 = phi [...], [ %IV1, %latch ]
3621 // In this case, if IV1 has an external use, we need to avoid adding both
3622 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3623 // don't already have an incoming value for the middle block.
3624 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3625 PHI->addIncoming(I.second, MiddleBlock);
3626 }
3627}
3628
3629namespace {
3630struct CSEDenseMapInfo {
3631 static bool canHandle(Instruction *I) {
3632 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3633 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3634 }
3635 static inline Instruction *getEmptyKey() {
3636 return DenseMapInfo<Instruction *>::getEmptyKey();
3637 }
3638 static inline Instruction *getTombstoneKey() {
3639 return DenseMapInfo<Instruction *>::getTombstoneKey();
3640 }
3641 static unsigned getHashValue(Instruction *I) {
3642 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3642, __PRETTY_FUNCTION__))
;
3643 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3644 I->value_op_end()));
3645 }
3646 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3647 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3648 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3649 return LHS == RHS;
3650 return LHS->isIdenticalTo(RHS);
3651 }
3652};
3653}
3654
3655///\brief Perform cse of induction variable instructions.
3656static void cse(BasicBlock *BB) {
3657 // Perform simple cse.
3658 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3659 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3660 Instruction *In = &*I++;
3661
3662 if (!CSEDenseMapInfo::canHandle(In))
3663 continue;
3664
3665 // Check if we can replace this instruction with any of the
3666 // visited instructions.
3667 if (Instruction *V = CSEMap.lookup(In)) {
3668 In->replaceAllUsesWith(V);
3669 In->eraseFromParent();
3670 continue;
3671 }
3672
3673 CSEMap[In] = In;
3674 }
3675}
3676
3677/// \brief Adds a 'fast' flag to floating point operations.
3678static Value *addFastMathFlag(Value *V) {
3679 if (isa<FPMathOperator>(V)) {
3680 FastMathFlags Flags;
3681 Flags.setUnsafeAlgebra();
3682 cast<Instruction>(V)->setFastMathFlags(Flags);
3683 }
3684 return V;
3685}
3686
3687/// \brief Estimate the overhead of scalarizing an instruction. This is a
3688/// convenience wrapper for the type-based getScalarizationOverhead API.
3689static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3690 const TargetTransformInfo &TTI) {
3691 if (VF == 1)
3692 return 0;
3693
3694 unsigned Cost = 0;
3695 Type *RetTy = ToVectorTy(I->getType(), VF);
3696 if (!RetTy->isVoidTy())
3697 Cost += TTI.getScalarizationOverhead(RetTy, true, false);
3698
3699 if (CallInst *CI = dyn_cast<CallInst>(I)) {
3700 SmallVector<const Value *, 4> Operands(CI->arg_operands());
3701 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3702 } else {
3703 SmallVector<const Value *, 4> Operands(I->operand_values());
3704 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3705 }
3706
3707 return Cost;
3708}
3709
3710// Estimate cost of a call instruction CI if it were vectorized with factor VF.
3711// Return the cost of the instruction, including scalarization overhead if it's
3712// needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3713// i.e. either vector version isn't available, or is too expensive.
3714static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3715 const TargetTransformInfo &TTI,
3716 const TargetLibraryInfo *TLI,
3717 bool &NeedToScalarize) {
3718 Function *F = CI->getCalledFunction();
3719 StringRef FnName = CI->getCalledFunction()->getName();
3720 Type *ScalarRetTy = CI->getType();
3721 SmallVector<Type *, 4> Tys, ScalarTys;
3722 for (auto &ArgOp : CI->arg_operands())
3723 ScalarTys.push_back(ArgOp->getType());
3724
3725 // Estimate cost of scalarized vector call. The source operands are assumed
3726 // to be vectors, so we need to extract individual elements from there,
3727 // execute VF scalar calls, and then gather the result into the vector return
3728 // value.
3729 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3730 if (VF == 1)
3731 return ScalarCallCost;
3732
3733 // Compute corresponding vector type for return value and arguments.
3734 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3735 for (Type *ScalarTy : ScalarTys)
3736 Tys.push_back(ToVectorTy(ScalarTy, VF));
3737
3738 // Compute costs of unpacking argument values for the scalar calls and
3739 // packing the return values to a vector.
3740 unsigned ScalarizationCost = getScalarizationOverhead(CI, VF, TTI);
3741
3742 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3743
3744 // If we can't emit a vector call for this function, then the currently found
3745 // cost is the cost we need to return.
3746 NeedToScalarize = true;
3747 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3748 return Cost;
3749
3750 // If the corresponding vector cost is cheaper, return its cost.
3751 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3752 if (VectorCallCost < Cost) {
3753 NeedToScalarize = false;
3754 return VectorCallCost;
3755 }
3756 return Cost;
3757}
3758
3759// Estimate cost of an intrinsic call instruction CI if it were vectorized with
3760// factor VF. Return the cost of the instruction, including scalarization
3761// overhead if it's needed.
3762static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3763 const TargetTransformInfo &TTI,
3764 const TargetLibraryInfo *TLI) {
3765 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3766 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3766, __PRETTY_FUNCTION__))
;
3767
3768 Type *RetTy = ToVectorTy(CI->getType(), VF);
3769 SmallVector<Type *, 4> Tys;
3770 for (Value *ArgOperand : CI->arg_operands())
3771 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
3772
3773 FastMathFlags FMF;
3774 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3775 FMF = FPMO->getFastMathFlags();
3776
3777 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys, FMF);
3778}
3779
3780static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3781 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3782 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3783 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3784}
3785static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3786 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3787 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3788 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3789}
3790
3791void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3792 // For every instruction `I` in MinBWs, truncate the operands, create a
3793 // truncated version of `I` and reextend its result. InstCombine runs
3794 // later and will remove any ext/trunc pairs.
3795 //
3796 SmallPtrSet<Value *, 4> Erased;
3797 for (const auto &KV : Cost->getMinimalBitwidths()) {
3798 // If the value wasn't vectorized, we must maintain the original scalar
3799 // type. The absence of the value from VectorLoopValueMap indicates that it
3800 // wasn't vectorized.
3801 if (!VectorLoopValueMap.hasVector(KV.first))
3802 continue;
3803 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3804 for (Value *&I : Parts) {
3805 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3806 continue;
3807 Type *OriginalTy = I->getType();
3808 Type *ScalarTruncatedTy =
3809 IntegerType::get(OriginalTy->getContext(), KV.second);
3810 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3811 OriginalTy->getVectorNumElements());
3812 if (TruncatedTy == OriginalTy)
3813 continue;
3814
3815 IRBuilder<> B(cast<Instruction>(I));
3816 auto ShrinkOperand = [&](Value *V) -> Value * {
3817 if (auto *ZI = dyn_cast<ZExtInst>(V))
3818 if (ZI->getSrcTy() == TruncatedTy)
3819 return ZI->getOperand(0);
3820 return B.CreateZExtOrTrunc(V, TruncatedTy);
3821 };
3822
3823 // The actual instruction modification depends on the instruction type,
3824 // unfortunately.
3825 Value *NewI = nullptr;
3826 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3827 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3828 ShrinkOperand(BO->getOperand(1)));
3829 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3830 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3831 NewI =
3832 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3833 ShrinkOperand(CI->getOperand(1)));
3834 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3835 NewI = B.CreateSelect(SI->getCondition(),
3836 ShrinkOperand(SI->getTrueValue()),
3837 ShrinkOperand(SI->getFalseValue()));
3838 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3839 switch (CI->getOpcode()) {
3840 default:
3841 llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3841)
;
3842 case Instruction::Trunc:
3843 NewI = ShrinkOperand(CI->getOperand(0));
3844 break;
3845 case Instruction::SExt:
3846 NewI = B.CreateSExtOrTrunc(
3847 CI->getOperand(0),
3848 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3849 break;
3850 case Instruction::ZExt:
3851 NewI = B.CreateZExtOrTrunc(
3852 CI->getOperand(0),
3853 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3854 break;
3855 }
3856 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3857 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3858 auto *O0 = B.CreateZExtOrTrunc(
3859 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3860 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3861 auto *O1 = B.CreateZExtOrTrunc(
3862 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3863
3864 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3865 } else if (isa<LoadInst>(I)) {
3866 // Don't do anything with the operands, just extend the result.
3867 continue;
3868 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3869 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3870 auto *O0 = B.CreateZExtOrTrunc(
3871 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3872 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3873 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3874 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3875 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3876 auto *O0 = B.CreateZExtOrTrunc(
3877 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3878 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3879 } else {
3880 llvm_unreachable("Unhandled instruction type!")::llvm::llvm_unreachable_internal("Unhandled instruction type!"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3880)
;
3881 }
3882
3883 // Lastly, extend the result.
3884 NewI->takeName(cast<Instruction>(I));
3885 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3886 I->replaceAllUsesWith(Res);
3887 cast<Instruction>(I)->eraseFromParent();
3888 Erased.insert(I);
3889 I = Res;
3890 }
3891 }
3892
3893 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3894 for (const auto &KV : Cost->getMinimalBitwidths()) {
3895 // If the value wasn't vectorized, we must maintain the original scalar
3896 // type. The absence of the value from VectorLoopValueMap indicates that it
3897 // wasn't vectorized.
3898 if (!VectorLoopValueMap.hasVector(KV.first))
3899 continue;
3900 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3901 for (Value *&I : Parts) {
3902 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3903 if (Inst && Inst->use_empty()) {
3904 Value *NewI = Inst->getOperand(0);
3905 Inst->eraseFromParent();
3906 I = NewI;
3907 }
3908 }
3909 }
3910}
3911
3912void InnerLoopVectorizer::vectorizeLoop() {
3913 //===------------------------------------------------===//
3914 //
3915 // Notice: any optimization or new instruction that go
3916 // into the code below should be also be implemented in
3917 // the cost-model.
3918 //
3919 //===------------------------------------------------===//
3920 Constant *Zero = Builder.getInt32(0);
3921
3922 // In order to support recurrences we need to be able to vectorize Phi nodes.
3923 // Phi nodes have cycles, so we need to vectorize them in two stages. First,
3924 // we create a new vector PHI node with no incoming edges. We use this value
3925 // when we vectorize all of the instructions that use the PHI. Next, after
3926 // all of the instructions in the block are complete we add the new incoming
3927 // edges to the PHI. At this point all of the instructions in the basic block
3928 // are vectorized, so we can use them to construct the PHI.
3929 PhiVector PHIsToFix;
3930
3931 // Collect instructions from the original loop that will become trivially
3932 // dead in the vectorized loop. We don't need to vectorize these
3933 // instructions.
3934 collectTriviallyDeadInstructions();
3935
3936 // Scan the loop in a topological order to ensure that defs are vectorized
3937 // before users.
3938 LoopBlocksDFS DFS(OrigLoop);
3939 DFS.perform(LI);
3940
3941 // Vectorize all of the blocks in the original loop.
3942 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
3943 vectorizeBlockInLoop(BB, &PHIsToFix);
3944
3945 // Insert truncates and extends for any truncated instructions as hints to
3946 // InstCombine.
3947 if (VF > 1)
3948 truncateToMinimalBitwidths();
3949
3950 // At this point every instruction in the original loop is widened to a
3951 // vector form. Now we need to fix the recurrences in PHIsToFix. These PHI
3952 // nodes are currently empty because we did not want to introduce cycles.
3953 // This is the second stage of vectorizing recurrences.
3954 for (PHINode *Phi : PHIsToFix) {
3955 assert(Phi && "Unable to recover vectorized PHI")((Phi && "Unable to recover vectorized PHI") ? static_cast
<void> (0) : __assert_fail ("Phi && \"Unable to recover vectorized PHI\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3955, __PRETTY_FUNCTION__))
;
3956
3957 // Handle first-order recurrences that need to be fixed.
3958 if (Legal->isFirstOrderRecurrence(Phi)) {
3959 fixFirstOrderRecurrence(Phi);
3960 continue;
3961 }
3962
3963 // If the phi node is not a first-order recurrence, it must be a reduction.
3964 // Get it's reduction variable descriptor.
3965 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3966, __PRETTY_FUNCTION__))
3966 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3966, __PRETTY_FUNCTION__))
;
3967 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
3968
3969 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3970 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3971 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3972 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3973 RdxDesc.getMinMaxRecurrenceKind();
3974 setDebugLocFromInst(Builder, ReductionStartValue);
3975
3976 // We need to generate a reduction vector from the incoming scalar.
3977 // To do so, we need to generate the 'identity' vector and override
3978 // one of the elements with the incoming scalar reduction. We need
3979 // to do it in the vector-loop preheader.
3980 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3981
3982 // This is the vector-clone of the value that leaves the loop.
3983 const VectorParts &VectorExit = getVectorValue(LoopExitInst);
3984 Type *VecTy = VectorExit[0]->getType();
3985
3986 // Find the reduction identity variable. Zero for addition, or, xor,
3987 // one for multiplication, -1 for And.
3988 Value *Identity;
3989 Value *VectorStart;
3990 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3991 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3992 // MinMax reduction have the start value as their identify.
3993 if (VF == 1) {
3994 VectorStart = Identity = ReductionStartValue;
3995 } else {
3996 VectorStart = Identity =
3997 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3998 }
3999 } else {
4000 // Handle other reduction kinds:
4001 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
4002 RK, VecTy->getScalarType());
4003 if (VF == 1) {
4004 Identity = Iden;
4005 // This vector is the Identity vector where the first element is the
4006 // incoming scalar reduction.
4007 VectorStart = ReductionStartValue;
4008 } else {
4009 Identity = ConstantVector::getSplat(VF, Iden);
4010
4011 // This vector is the Identity vector where the first element is the
4012 // incoming scalar reduction.
4013 VectorStart =
4014 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
4015 }
4016 }
4017
4018 // Fix the vector-loop phi.
4019
4020 // Reductions do not have to start at zero. They can start with
4021 // any loop invariant values.
4022 const VectorParts &VecRdxPhi = getVectorValue(Phi);
4023 BasicBlock *Latch = OrigLoop->getLoopLatch();
4024 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
4025 const VectorParts &Val = getVectorValue(LoopVal);
4026 for (unsigned part = 0; part < UF; ++part) {
4027 // Make sure to add the reduction stat value only to the
4028 // first unroll part.
4029 Value *StartVal = (part == 0) ? VectorStart : Identity;
4030 cast<PHINode>(VecRdxPhi[part])
4031 ->addIncoming(StartVal, LoopVectorPreHeader);
4032 cast<PHINode>(VecRdxPhi[part])
4033 ->addIncoming(Val[part], LoopVectorBody);
4034 }
4035
4036 // Before each round, move the insertion point right between
4037 // the PHIs and the values we are going to write.
4038 // This allows us to write both PHINodes and the extractelement
4039 // instructions.
4040 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4041
4042 VectorParts &RdxParts = VectorLoopValueMap.getVector(LoopExitInst);
4043 setDebugLocFromInst(Builder, LoopExitInst);
4044
4045 // If the vector reduction can be performed in a smaller type, we truncate
4046 // then extend the loop exit value to enable InstCombine to evaluate the
4047 // entire expression in the smaller type.
4048 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
4049 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
4050 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
4051 for (unsigned part = 0; part < UF; ++part) {
4052 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4053 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4054 : Builder.CreateZExt(Trunc, VecTy);
4055 for (Value::user_iterator UI = RdxParts[part]->user_begin();
4056 UI != RdxParts[part]->user_end();)
4057 if (*UI != Trunc) {
4058 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
4059 RdxParts[part] = Extnd;
4060 } else {
4061 ++UI;
4062 }
4063 }
4064 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4065 for (unsigned part = 0; part < UF; ++part)
4066 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4067 }
4068
4069 // Reduce all of the unrolled parts into a single vector.
4070 Value *ReducedPartRdx = RdxParts[0];
4071 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
4072 setDebugLocFromInst(Builder, ReducedPartRdx);
4073 for (unsigned part = 1; part < UF; ++part) {
4074 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4075 // Floating point operations had to be 'fast' to enable the reduction.
4076 ReducedPartRdx = addFastMathFlag(
4077 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
4078 ReducedPartRdx, "bin.rdx"));
4079 else
4080 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4081 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
4082 }
4083
4084 if (VF > 1) {
4085 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
4086 // and vector ops, reducing the set of values being computed by half each
4087 // round.
4088 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4089, __PRETTY_FUNCTION__))
4089 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4089, __PRETTY_FUNCTION__))
;
4090 Value *TmpVec = ReducedPartRdx;
4091 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
4092 for (unsigned i = VF; i != 1; i >>= 1) {
4093 // Move the upper half of the vector to the lower half.
4094 for (unsigned j = 0; j != i / 2; ++j)
4095 ShuffleMask[j] = Builder.getInt32(i / 2 + j);
4096
4097 // Fill the rest of the mask with undef.
4098 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
4099 UndefValue::get(Builder.getInt32Ty()));
4100
4101 Value *Shuf = Builder.CreateShuffleVector(
4102 TmpVec, UndefValue::get(TmpVec->getType()),
4103 ConstantVector::get(ShuffleMask), "rdx.shuf");
4104
4105 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4106 // Floating point operations had to be 'fast' to enable the reduction.
4107 TmpVec = addFastMathFlag(Builder.CreateBinOp(
4108 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
4109 else
4110 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
4111 TmpVec, Shuf);
4112 }
4113
4114 // The result is in the first element of the vector.
4115 ReducedPartRdx =
4116 Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4117
4118 // If the reduction can be performed in a smaller type, we need to extend
4119 // the reduction to the wider type before we branch to the original loop.
4120 if (Phi->getType() != RdxDesc.getRecurrenceType())
4121 ReducedPartRdx =
4122 RdxDesc.isSigned()
4123 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4124 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4125 }
4126
4127 // Create a phi node that merges control-flow from the backedge-taken check
4128 // block and the middle block.
4129 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4130 LoopScalarPreHeader->getTerminator());
4131 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4132 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4133 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4134
4135 // Now, we need to fix the users of the reduction variable
4136 // inside and outside of the scalar remainder loop.
4137 // We know that the loop is in LCSSA form. We need to update the
4138 // PHI nodes in the exit blocks.
4139 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4140 LEE = LoopExitBlock->end();
4141 LEI != LEE; ++LEI) {
4142 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4143 if (!LCSSAPhi)
4144 break;
4145
4146 // All PHINodes need to have a single entry edge, or two if
4147 // we already fixed them.
4148 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4148, __PRETTY_FUNCTION__))
;
4149
4150 // We found a reduction value exit-PHI. Update it with the
4151 // incoming bypass edge.
4152 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst)
4153 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4154 } // end of the LCSSA phi scan.
4155
4156 // Fix the scalar loop reduction variable with the incoming reduction sum
4157 // from the vector body and from the backedge value.
4158 int IncomingEdgeBlockIdx =
4159 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
4160 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4160, __PRETTY_FUNCTION__))
;
4161 // Pick the other block.
4162 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4163 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4164 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4165 } // end of for each Phi in PHIsToFix.
4166
4167 // Update the dominator tree.
4168 //
4169 // FIXME: After creating the structure of the new loop, the dominator tree is
4170 // no longer up-to-date, and it remains that way until we update it
4171 // here. An out-of-date dominator tree is problematic for SCEV,
4172 // because SCEVExpander uses it to guide code generation. The
4173 // vectorizer use SCEVExpanders in several places. Instead, we should
4174 // keep the dominator tree up-to-date as we go.
4175 updateAnalysis();
4176
4177 // Fix-up external users of the induction variables.
4178 for (auto &Entry : *Legal->getInductionVars())
4179 fixupIVUsers(Entry.first, Entry.second,
4180 getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)),
4181 IVEndValues[Entry.first], LoopMiddleBlock);
4182
4183 fixLCSSAPHIs();
4184 predicateInstructions();
4185
4186 // Remove redundant induction instructions.
4187 cse(LoopVectorBody);
4188}
4189
4190void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
4191
4192 // This is the second phase of vectorizing first-order recurrences. An
4193 // overview of the transformation is described below. Suppose we have the
4194 // following loop.
4195 //
4196 // for (int i = 0; i < n; ++i)
4197 // b[i] = a[i] - a[i - 1];
4198 //
4199 // There is a first-order recurrence on "a". For this loop, the shorthand
4200 // scalar IR looks like:
4201 //
4202 // scalar.ph:
4203 // s_init = a[-1]
4204 // br scalar.body
4205 //
4206 // scalar.body:
4207 // i = phi [0, scalar.ph], [i+1, scalar.body]
4208 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
4209 // s2 = a[i]
4210 // b[i] = s2 - s1
4211 // br cond, scalar.body, ...
4212 //
4213 // In this example, s1 is a recurrence because it's value depends on the
4214 // previous iteration. In the first phase of vectorization, we created a
4215 // temporary value for s1. We now complete the vectorization and produce the
4216 // shorthand vector IR shown below (for VF = 4, UF = 1).
4217 //
4218 // vector.ph:
4219 // v_init = vector(..., ..., ..., a[-1])
4220 // br vector.body
4221 //
4222 // vector.body
4223 // i = phi [0, vector.ph], [i+4, vector.body]
4224 // v1 = phi [v_init, vector.ph], [v2, vector.body]
4225 // v2 = a[i, i+1, i+2, i+3];
4226 // v3 = vector(v1(3), v2(0, 1, 2))
4227 // b[i, i+1, i+2, i+3] = v2 - v3
4228 // br cond, vector.body, middle.block
4229 //
4230 // middle.block:
4231 // x = v2(3)
4232 // br scalar.ph
4233 //
4234 // scalar.ph:
4235 // s_init = phi [x, middle.block], [a[-1], otherwise]
4236 // br scalar.body
4237 //
4238 // After execution completes the vector loop, we extract the next value of
4239 // the recurrence (x) to use as the initial value in the scalar loop.
4240
4241 // Get the original loop preheader and single loop latch.
4242 auto *Preheader = OrigLoop->getLoopPreheader();
4243 auto *Latch = OrigLoop->getLoopLatch();
4244
4245 // Get the initial and previous values of the scalar recurrence.
4246 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4247 auto *Previous = Phi->getIncomingValueForBlock(Latch);
4248
4249 // Create a vector from the initial value.
4250 auto *VectorInit = ScalarInit;
4251 if (VF > 1) {
4252 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
4253 VectorInit = Builder.CreateInsertElement(
4254 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4255 Builder.getInt32(VF - 1), "vector.recur.init");
4256 }
4257
4258 // We constructed a temporary phi node in the first phase of vectorization.
4259 // This phi node will eventually be deleted.
4260 VectorParts &PhiParts = VectorLoopValueMap.getVector(Phi);
4261 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
4262
4263 // Create a phi node for the new recurrence. The current value will either be
4264 // the initial value inserted into a vector or loop-varying vector value.
4265 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4266 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4267
4268 // Get the vectorized previous value. We ensured the previous values was an
4269 // instruction when detecting the recurrence.
4270 auto &PreviousParts = getVectorValue(Previous);
4271
4272 // Set the insertion point to be after this instruction. We ensured the
4273 // previous value dominated all uses of the phi when detecting the
4274 // recurrence.
4275 Builder.SetInsertPoint(
4276 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
4277
4278 // We will construct a vector for the recurrence by combining the values for
4279 // the current and previous iterations. This is the required shuffle mask.
4280 SmallVector<Constant *, 8> ShuffleMask(VF);
4281 ShuffleMask[0] = Builder.getInt32(VF - 1);
4282 for (unsigned I = 1; I < VF; ++I)
4283 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4284
4285 // The vector from which to take the initial value for the current iteration
4286 // (actual or unrolled). Initially, this is the vector phi node.
4287 Value *Incoming = VecPhi;
4288
4289 // Shuffle the current and previous vector and update the vector parts.
4290 for (unsigned Part = 0; Part < UF; ++Part) {
4291 auto *Shuffle =
4292 VF > 1
4293 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
4294 ConstantVector::get(ShuffleMask))
4295 : Incoming;
4296 PhiParts[Part]->replaceAllUsesWith(Shuffle);
4297 cast<Instruction>(PhiParts[Part])->eraseFromParent();
4298 PhiParts[Part] = Shuffle;
4299 Incoming = PreviousParts[Part];
4300 }
4301
4302 // Fix the latch value of the new recurrence in the vector loop.
4303 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4304
4305 // Extract the last vector element in the middle block. This will be the
4306 // initial value for the recurrence when jumping to the scalar loop.
4307 auto *Extract = Incoming;
4308 if (VF > 1) {
4309 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
4310 Extract = Builder.CreateExtractElement(Extract, Builder.getInt32(VF - 1),
4311 "vector.recur.extract");
4312 }
4313
4314 // Fix the initial value of the original recurrence in the scalar loop.
4315 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
4316 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4317 for (auto *BB : predecessors(LoopScalarPreHeader)) {
4318 auto *Incoming = BB == LoopMiddleBlock ? Extract : ScalarInit;
4319 Start->addIncoming(Incoming, BB);
4320 }
4321
4322 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
4323 Phi->setName("scalar.recur");
4324
4325 // Finally, fix users of the recurrence outside the loop. The users will need
4326 // either the last value of the scalar recurrence or the last value of the
4327 // vector recurrence we extracted in the middle block. Since the loop is in
4328 // LCSSA form, we just need to find the phi node for the original scalar
4329 // recurrence in the exit block, and then add an edge for the middle block.
4330 for (auto &I : *LoopExitBlock) {
4331 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4332 if (!LCSSAPhi)
4333 break;
4334 if (LCSSAPhi->getIncomingValue(0) == Phi) {
4335 LCSSAPhi->addIncoming(Extract, LoopMiddleBlock);
4336 break;
4337 }
4338 }
4339}
4340
4341void InnerLoopVectorizer::fixLCSSAPHIs() {
4342 for (Instruction &LEI : *LoopExitBlock) {
4343 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4344 if (!LCSSAPhi)
4345 break;
4346 if (LCSSAPhi->getNumIncomingValues() == 1)
4347 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
4348 LoopMiddleBlock);
4349 }
4350}
4351
4352void InnerLoopVectorizer::collectTriviallyDeadInstructions() {
4353 BasicBlock *Latch = OrigLoop->getLoopLatch();
4354
4355 // We create new control-flow for the vectorized loop, so the original
4356 // condition will be dead after vectorization if it's only used by the
4357 // branch.
4358 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
4359 if (Cmp && Cmp->hasOneUse())
4360 DeadInstructions.insert(Cmp);
4361
4362 // We create new "steps" for induction variable updates to which the original
4363 // induction variables map. An original update instruction will be dead if
4364 // all its users except the induction variable are dead.
4365 for (auto &Induction : *Legal->getInductionVars()) {
4366 PHINode *Ind = Induction.first;
4367 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
4368 if (all_of(IndUpdate->users(), [&](User *U) -> bool {
4369 return U == Ind || DeadInstructions.count(cast<Instruction>(U));
4370 }))
4371 DeadInstructions.insert(IndUpdate);
4372 }
4373}
4374
4375void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) {
4376
4377 // The basic block and loop containing the predicated instruction.
4378 auto *PredBB = PredInst->getParent();
4379 auto *VectorLoop = LI->getLoopFor(PredBB);
4380
4381 // Initialize a worklist with the operands of the predicated instruction.
4382 SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4383
4384 // Holds instructions that we need to analyze again. An instruction may be
4385 // reanalyzed if we don't yet know if we can sink it or not.
4386 SmallVector<Instruction *, 8> InstsToReanalyze;
4387
4388 // Returns true if a given use occurs in the predicated block. Phi nodes use
4389 // their operands in their corresponding predecessor blocks.
4390 auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4391 auto *I = cast<Instruction>(U.getUser());
4392 BasicBlock *BB = I->getParent();
4393 if (auto *Phi = dyn_cast<PHINode>(I))
4394 BB = Phi->getIncomingBlock(
4395 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4396 return BB == PredBB;
4397 };
4398
4399 // Iteratively sink the scalarized operands of the predicated instruction
4400 // into the block we created for it. When an instruction is sunk, it's
4401 // operands are then added to the worklist. The algorithm ends after one pass
4402 // through the worklist doesn't sink a single instruction.
4403 bool Changed;
4404 do {
4405
4406 // Add the instructions that need to be reanalyzed to the worklist, and
4407 // reset the changed indicator.
4408 Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4409 InstsToReanalyze.clear();
4410 Changed = false;
4411
4412 while (!Worklist.empty()) {
4413 auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4414
4415 // We can't sink an instruction if it is a phi node, is already in the
4416 // predicated block, is not in the loop, or may have side effects.
4417 if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4418 !VectorLoop->contains(I) || I->mayHaveSideEffects())
4419 continue;
4420
4421 // It's legal to sink the instruction if all its uses occur in the
4422 // predicated block. Otherwise, there's nothing to do yet, and we may
4423 // need to reanalyze the instruction.
4424 if (!all_of(I->uses(), isBlockOfUsePredicated)) {
4425 InstsToReanalyze.push_back(I);
4426 continue;
4427 }
4428
4429 // Move the instruction to the beginning of the predicated block, and add
4430 // it's operands to the worklist.
4431 I->moveBefore(&*PredBB->getFirstInsertionPt());
4432 Worklist.insert(I->op_begin(), I->op_end());
4433
4434 // The sinking may have enabled other instructions to be sunk, so we will
4435 // need to iterate.
4436 Changed = true;
4437 }
4438 } while (Changed);
4439}
4440
4441void InnerLoopVectorizer::predicateInstructions() {
4442
4443 // For each instruction I marked for predication on value C, split I into its
4444 // own basic block to form an if-then construct over C. Since I may be fed by
4445 // an extractelement instruction or other scalar operand, we try to
4446 // iteratively sink its scalar operands into the predicated block. If I feeds
4447 // an insertelement instruction, we try to move this instruction into the
4448 // predicated block as well. For non-void types, a phi node will be created
4449 // for the resulting value (either vector or scalar).
4450 //
4451 // So for some predicated instruction, e.g. the conditional sdiv in:
4452 //
4453 // for.body:
4454 // ...
4455 // %add = add nsw i32 %mul, %0
4456 // %cmp5 = icmp sgt i32 %2, 7
4457 // br i1 %cmp5, label %if.then, label %if.end
4458 //
4459 // if.then:
4460 // %div = sdiv i32 %0, %1
4461 // br label %if.end
4462 //
4463 // if.end:
4464 // %x.0 = phi i32 [ %div, %if.then ], [ %add, %for.body ]
4465 //
4466 // the sdiv at this point is scalarized and if-converted using a select.
4467 // The inactive elements in the vector are not used, but the predicated
4468 // instruction is still executed for all vector elements, essentially:
4469 //
4470 // vector.body:
4471 // ...
4472 // %17 = add nsw <2 x i32> %16, %wide.load
4473 // %29 = extractelement <2 x i32> %wide.load, i32 0
4474 // %30 = extractelement <2 x i32> %wide.load51, i32 0
4475 // %31 = sdiv i32 %29, %30
4476 // %32 = insertelement <2 x i32> undef, i32 %31, i32 0
4477 // %35 = extractelement <2 x i32> %wide.load, i32 1
4478 // %36 = extractelement <2 x i32> %wide.load51, i32 1
4479 // %37 = sdiv i32 %35, %36
4480 // %38 = insertelement <2 x i32> %32, i32 %37, i32 1
4481 // %predphi = select <2 x i1> %26, <2 x i32> %38, <2 x i32> %17
4482 //
4483 // Predication will now re-introduce the original control flow to avoid false
4484 // side-effects by the sdiv instructions on the inactive elements, yielding
4485 // (after cleanup):
4486 //
4487 // vector.body:
4488 // ...
4489 // %5 = add nsw <2 x i32> %4, %wide.load
4490 // %8 = icmp sgt <2 x i32> %wide.load52, <i32 7, i32 7>
4491 // %9 = extractelement <2 x i1> %8, i32 0
4492 // br i1 %9, label %pred.sdiv.if, label %pred.sdiv.continue
4493 //
4494 // pred.sdiv.if:
4495 // %10 = extractelement <2 x i32> %wide.load, i32 0
4496 // %11 = extractelement <2 x i32> %wide.load51, i32 0
4497 // %12 = sdiv i32 %10, %11
4498 // %13 = insertelement <2 x i32> undef, i32 %12, i32 0
4499 // br label %pred.sdiv.continue
4500 //
4501 // pred.sdiv.continue:
4502 // %14 = phi <2 x i32> [ undef, %vector.body ], [ %13, %pred.sdiv.if ]
4503 // %15 = extractelement <2 x i1> %8, i32 1
4504 // br i1 %15, label %pred.sdiv.if54, label %pred.sdiv.continue55
4505 //
4506 // pred.sdiv.if54:
4507 // %16 = extractelement <2 x i32> %wide.load, i32 1
4508 // %17 = extractelement <2 x i32> %wide.load51, i32 1
4509 // %18 = sdiv i32 %16, %17
4510 // %19 = insertelement <2 x i32> %14, i32 %18, i32 1
4511 // br label %pred.sdiv.continue55
4512 //
4513 // pred.sdiv.continue55:
4514 // %20 = phi <2 x i32> [ %14, %pred.sdiv.continue ], [ %19, %pred.sdiv.if54 ]
4515 // %predphi = select <2 x i1> %8, <2 x i32> %20, <2 x i32> %5
4516
4517 for (auto KV : PredicatedInstructions) {
4518 BasicBlock::iterator I(KV.first);
4519 BasicBlock *Head = I->getParent();
4520 auto *BB = SplitBlock(Head, &*std::next(I), DT, LI);
4521 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
4522 /*BranchWeights=*/nullptr, DT, LI);
4523 I->moveBefore(T);
4524 sinkScalarOperands(&*I);
4525
4526 I->getParent()->setName(Twine("pred.") + I->getOpcodeName() + ".if");
4527 BB->setName(Twine("pred.") + I->getOpcodeName() + ".continue");
4528
4529 // If the instruction is non-void create a Phi node at reconvergence point.
4530 if (!I->getType()->isVoidTy()) {
4531 Value *IncomingTrue = nullptr;
4532 Value *IncomingFalse = nullptr;
4533
4534 if (I->hasOneUse() && isa<InsertElementInst>(*I->user_begin())) {
4535 // If the predicated instruction is feeding an insert-element, move it
4536 // into the Then block; Phi node will be created for the vector.
4537 InsertElementInst *IEI = cast<InsertElementInst>(*I->user_begin());
4538 IEI->moveBefore(T);
4539 IncomingTrue = IEI; // the new vector with the inserted element.
4540 IncomingFalse = IEI->getOperand(0); // the unmodified vector
4541 } else {
4542 // Phi node will be created for the scalar predicated instruction.
4543 IncomingTrue = &*I;
4544 IncomingFalse = UndefValue::get(I->getType());
4545 }
4546
4547 BasicBlock *PostDom = I->getParent()->getSingleSuccessor();
4548 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4548, __PRETTY_FUNCTION__))
;
4549 PHINode *Phi =
4550 PHINode::Create(IncomingTrue->getType(), 2, "", &PostDom->front());
4551 IncomingTrue->replaceAllUsesWith(Phi);
4552 Phi->addIncoming(IncomingFalse, Head);
4553 Phi->addIncoming(IncomingTrue, I->getParent());
4554 }
4555 }
4556
4557 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (false
)
;
4558}
4559
4560InnerLoopVectorizer::VectorParts
4561InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
4562 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4562, __PRETTY_FUNCTION__))
;
4563
4564 // Look for cached value.
4565 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
4566 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
4567 if (ECEntryIt != MaskCache.end())
4568 return ECEntryIt->second;
4569
4570 VectorParts SrcMask = createBlockInMask(Src);
4571
4572 // The terminator has to be a branch inst!
4573 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
4574 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4574, __PRETTY_FUNCTION__))
;
4575
4576 if (BI->isConditional()) {
4577 VectorParts EdgeMask = getVectorValue(BI->getCondition());
4578
4579 if (BI->getSuccessor(0) != Dst)
4580 for (unsigned part = 0; part < UF; ++part)
4581 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4582
4583 for (unsigned part = 0; part < UF; ++part)
4584 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4585
4586 MaskCache[Edge] = EdgeMask;
4587 return EdgeMask;
4588 }
4589
4590 MaskCache[Edge] = SrcMask;
4591 return SrcMask;
4592}
4593
4594InnerLoopVectorizer::VectorParts
4595InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4596 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4596, __PRETTY_FUNCTION__))
;
4597
4598 // Loop incoming mask is all-one.
4599 if (OrigLoop->getHeader() == BB) {
4600 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4601 return getVectorValue(C);
4602 }
4603
4604 // This is the block mask. We OR all incoming edges, and with zero.
4605 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4606 VectorParts BlockMask = getVectorValue(Zero);
4607
4608 // For each pred:
4609 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4610 VectorParts EM = createEdgeMask(*it, BB);
4611 for (unsigned part = 0; part < UF; ++part)
4612 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4613 }
4614
4615 return BlockMask;
4616}
4617
4618void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
4619 unsigned VF, PhiVector *PV) {
4620 PHINode *P = cast<PHINode>(PN);
4621 // Handle recurrences.
4622 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4623 VectorParts Entry(UF);
4624 for (unsigned part = 0; part < UF; ++part) {
4625 // This is phase one of vectorizing PHIs.
4626 Type *VecTy =
4627 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4628 Entry[part] = PHINode::Create(
4629 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4630 }
4631 VectorLoopValueMap.initVector(P, Entry);
4632 PV->push_back(P);
4633 return;
4634 }
4635
4636 setDebugLocFromInst(Builder, P);
4637 // Check for PHI nodes that are lowered to vector selects.
4638 if (P->getParent() != OrigLoop->getHeader()) {
4639 // We know that all PHIs in non-header blocks are converted into
4640 // selects, so we don't have to worry about the insertion order and we
4641 // can just use the builder.
4642 // At this point we generate the predication tree. There may be
4643 // duplications since this is a simple recursive scan, but future
4644 // optimizations will clean it up.
4645
4646 unsigned NumIncoming = P->getNumIncomingValues();
4647
4648 // Generate a sequence of selects of the form:
4649 // SELECT(Mask3, In3,
4650 // SELECT(Mask2, In2,
4651 // ( ...)))
4652 VectorParts Entry(UF);
4653 for (unsigned In = 0; In < NumIncoming; In++) {
4654 VectorParts Cond =
4655 createEdgeMask(P->getIncomingBlock(In), P->getParent());
4656 const VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
4657
4658 for (unsigned part = 0; part < UF; ++part) {
4659 // We might have single edge PHIs (blocks) - use an identity
4660 // 'select' for the first PHI operand.
4661 if (In == 0)
4662 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In0[part]);
4663 else
4664 // Select between the current value and the previous incoming edge
4665 // based on the incoming mask.
4666 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], Entry[part],
4667 "predphi");
4668 }
4669 }
4670 VectorLoopValueMap.initVector(P, Entry);
4671 return;
4672 }
4673
4674 // This PHINode must be an induction variable.
4675 // Make sure that we know about it.
4676 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4676, __PRETTY_FUNCTION__))
;
4677
4678 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4679 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4680
4681 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4682 // which can be found from the original scalar operations.
4683 switch (II.getKind()) {
4684 case InductionDescriptor::IK_NoInduction:
4685 llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4685)
;
4686 case InductionDescriptor::IK_IntInduction:
4687 return widenIntInduction(P);
4688 case InductionDescriptor::IK_PtrInduction: {
4689 // Handle the pointer induction variable case.
4690 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4690, __PRETTY_FUNCTION__))
;
4691 // This is the normalized GEP that starts counting at zero.
4692 Value *PtrInd = Induction;
4693 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4694 // Determine the number of scalars we need to generate for each unroll
4695 // iteration. If the instruction is uniform, we only need to generate the
4696 // first lane. Otherwise, we generate all VF values.
4697 unsigned Lanes = Cost->isUniformAfterVectorization(P, VF) ? 1 : VF;
4698 // These are the scalar results. Notice that we don't generate vector GEPs
4699 // because scalar GEPs result in better code.
4700 ScalarParts Entry(UF);
4701 for (unsigned Part = 0; Part < UF; ++Part) {
4702 Entry[Part].resize(VF);
4703 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4704 Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4705 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4706 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4707 SclrGep->setName("next.gep");
4708 Entry[Part][Lane] = SclrGep;
4709 }
4710 }
4711 VectorLoopValueMap.initScalar(P, Entry);
4712 return;
4713 }
4714 case InductionDescriptor::IK_FpInduction: {
4715 assert(P->getType() == II.getStartValue()->getType() &&((P->getType() == II.getStartValue()->getType() &&
"Types must match") ? static_cast<void> (0) : __assert_fail
("P->getType() == II.getStartValue()->getType() && \"Types must match\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4716, __PRETTY_FUNCTION__))
4716 "Types must match")((P->getType() == II.getStartValue()->getType() &&
"Types must match") ? static_cast<void> (0) : __assert_fail
("P->getType() == II.getStartValue()->getType() && \"Types must match\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4716, __PRETTY_FUNCTION__))
;
4717 // Handle other induction variables that are now based on the
4718 // canonical one.
4719 assert(P != OldInduction && "Primary induction can be integer only")((P != OldInduction && "Primary induction can be integer only"
) ? static_cast<void> (0) : __assert_fail ("P != OldInduction && \"Primary induction can be integer only\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4719, __PRETTY_FUNCTION__))
;
4720
4721 Value *V = Builder.CreateCast(Instruction::SIToFP, Induction, P->getType());
4722 V = II.transform(Builder, V, PSE.getSE(), DL);
4723 V->setName("fp.offset.idx");
4724
4725 // Now we have scalar op: %fp.offset.idx = StartVal +/- Induction*StepVal
4726
4727 Value *Broadcasted = getBroadcastInstrs(V);
4728 // After broadcasting the induction variable we need to make the vector
4729 // consecutive by adding StepVal*0, StepVal*1, StepVal*2, etc.
4730 Value *StepVal = cast<SCEVUnknown>(II.getStep())->getValue();
4731 VectorParts Entry(UF);
4732 for (unsigned part = 0; part < UF; ++part)
4733 Entry[part] = getStepVector(Broadcasted, VF * part, StepVal,
4734 II.getInductionOpcode());
4735 VectorLoopValueMap.initVector(P, Entry);
4736 return;
4737 }
4738 }
4739}
4740
4741/// A helper function for checking whether an integer division-related
4742/// instruction may divide by zero (in which case it must be predicated if
4743/// executed conditionally in the scalar code).
4744/// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4745/// Non-zero divisors that are non compile-time constants will not be
4746/// converted into multiplication, so we will still end up scalarizing
4747/// the division, but can do so w/o predication.
4748static bool mayDivideByZero(Instruction &I) {
4749 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4753, __PRETTY_FUNCTION__))
4750 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4753, __PRETTY_FUNCTION__))
4751 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4753, __PRETTY_FUNCTION__))
4752 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4753, __PRETTY_FUNCTION__))
4753 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4753, __PRETTY_FUNCTION__))
;
4754 Value *Divisor = I.getOperand(1);
4755 auto *CInt = dyn_cast<ConstantInt>(Divisor);
4756 return !CInt || CInt->isZero();
4757}
4758
4759void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
4760 // For each instruction in the old loop.
4761 for (Instruction &I : *BB) {
4762
4763 // If the instruction will become trivially dead when vectorized, we don't
4764 // need to generate it.
4765 if (DeadInstructions.count(&I))
4766 continue;
4767
4768 // Scalarize instructions that should remain scalar after vectorization.
4769 if (VF > 1 &&
4770 !(isa<BranchInst>(&I) || isa<PHINode>(&I) ||
4771 isa<DbgInfoIntrinsic>(&I)) &&
4772 shouldScalarizeInstruction(&I)) {
4773 scalarizeInstruction(&I, Legal->isScalarWithPredication(&I));
4774 continue;
4775 }
4776
4777 switch (I.getOpcode()) {
4778 case Instruction::Br:
4779 // Nothing to do for PHIs and BR, since we already took care of the
4780 // loop control flow instructions.
4781 continue;
4782 case Instruction::PHI: {
4783 // Vectorize PHINodes.
4784 widenPHIInstruction(&I, UF, VF, PV);
4785 continue;
4786 } // End of PHI.
4787
4788 case Instruction::UDiv:
4789 case Instruction::SDiv:
4790 case Instruction::SRem:
4791 case Instruction::URem:
4792 // Scalarize with predication if this instruction may divide by zero and
4793 // block execution is conditional, otherwise fallthrough.
4794 if (Legal->isScalarWithPredication(&I)) {
4795 scalarizeInstruction(&I, true);
4796 continue;
4797 }
4798 case Instruction::Add:
4799 case Instruction::FAdd:
4800 case Instruction::Sub:
4801 case Instruction::FSub:
4802 case Instruction::Mul:
4803 case Instruction::FMul:
4804 case Instruction::FDiv:
4805 case Instruction::FRem:
4806 case Instruction::Shl:
4807 case Instruction::LShr:
4808 case Instruction::AShr:
4809 case Instruction::And:
4810 case Instruction::Or:
4811 case Instruction::Xor: {
4812 // Just widen binops.
4813 auto *BinOp = cast<BinaryOperator>(&I);
4814 setDebugLocFromInst(Builder, BinOp);
4815 const VectorParts &A = getVectorValue(BinOp->getOperand(0));
4816 const VectorParts &B = getVectorValue(BinOp->getOperand(1));
4817
4818 // Use this vector value for all users of the original instruction.
4819 VectorParts Entry(UF);
4820 for (unsigned Part = 0; Part < UF; ++Part) {
4821 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4822
4823 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4824 VecOp->copyIRFlags(BinOp);
4825
4826 Entry[Part] = V;
4827 }
4828
4829 VectorLoopValueMap.initVector(&I, Entry);
4830 addMetadata(Entry, BinOp);
4831 break;
4832 }
4833 case Instruction::Select: {
4834 // Widen selects.
4835 // If the selector is loop invariant we can create a select
4836 // instruction with a scalar condition. Otherwise, use vector-select.
4837 auto *SE = PSE.getSE();
4838 bool InvariantCond =
4839 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4840 setDebugLocFromInst(Builder, &I);
4841
4842 // The condition can be loop invariant but still defined inside the
4843 // loop. This means that we can't just use the original 'cond' value.
4844 // We have to take the 'vectorized' value and pick the first lane.
4845 // Instcombine will make this a no-op.
4846 const VectorParts &Cond = getVectorValue(I.getOperand(0));
4847 const VectorParts &Op0 = getVectorValue(I.getOperand(1));
4848 const VectorParts &Op1 = getVectorValue(I.getOperand(2));
4849
4850 auto *ScalarCond = getScalarValue(I.getOperand(0), 0, 0);
4851
4852 VectorParts Entry(UF);
4853 for (unsigned Part = 0; Part < UF; ++Part) {
4854 Entry[Part] = Builder.CreateSelect(
4855 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4856 }
4857
4858 VectorLoopValueMap.initVector(&I, Entry);
4859 addMetadata(Entry, &I);
4860 break;
4861 }
4862
4863 case Instruction::ICmp:
4864 case Instruction::FCmp: {
4865 // Widen compares. Generate vector compares.
4866 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4867 auto *Cmp = dyn_cast<CmpInst>(&I);
4868 setDebugLocFromInst(Builder, Cmp);
4869 const VectorParts &A = getVectorValue(Cmp->getOperand(0));
4870 const VectorParts &B = getVectorValue(Cmp->getOperand(1));
4871 VectorParts Entry(UF);
4872 for (unsigned Part = 0; Part < UF; ++Part) {
4873 Value *C = nullptr;
4874 if (FCmp) {
4875 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4876 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4877 } else {
4878 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4879 }
4880 Entry[Part] = C;
4881 }
4882
4883 VectorLoopValueMap.initVector(&I, Entry);
4884 addMetadata(Entry, &I);
4885 break;
4886 }
4887
4888 case Instruction::Store:
4889 case Instruction::Load:
4890 vectorizeMemoryInstruction(&I);
4891 break;
4892 case Instruction::ZExt:
4893 case Instruction::SExt:
4894 case Instruction::FPToUI:
4895 case Instruction::FPToSI:
4896 case Instruction::FPExt:
4897 case Instruction::PtrToInt:
4898 case Instruction::IntToPtr:
4899 case Instruction::SIToFP:
4900 case Instruction::UIToFP:
4901 case Instruction::Trunc:
4902 case Instruction::FPTrunc:
4903 case Instruction::BitCast: {
4904 auto *CI = dyn_cast<CastInst>(&I);
4905 setDebugLocFromInst(Builder, CI);
4906
4907 // Optimize the special case where the source is a constant integer
4908 // induction variable. Notice that we can only optimize the 'trunc' case
4909 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4910 // (c) other casts depend on pointer size.
4911 if (Cost->isOptimizableIVTruncate(CI, VF)) {
4912 widenIntInduction(cast<PHINode>(CI->getOperand(0)),
4913 cast<TruncInst>(CI));
4914 break;
4915 }
4916
4917 /// Vectorize casts.
4918 Type *DestTy =
4919 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4920
4921 const VectorParts &A = getVectorValue(CI->getOperand(0));
4922 VectorParts Entry(UF);
4923 for (unsigned Part = 0; Part < UF; ++Part)
4924 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4925 VectorLoopValueMap.initVector(&I, Entry);
4926 addMetadata(Entry, &I);
4927 break;
4928 }
4929
4930 case Instruction::Call: {
4931 // Ignore dbg intrinsics.
4932 if (isa<DbgInfoIntrinsic>(I))
4933 break;
4934 setDebugLocFromInst(Builder, &I);
4935
4936 Module *M = BB->getParent()->getParent();
4937 auto *CI = cast<CallInst>(&I);
4938
4939 StringRef FnName = CI->getCalledFunction()->getName();
4940 Function *F = CI->getCalledFunction();
4941 Type *RetTy = ToVectorTy(CI->getType(), VF);
4942 SmallVector<Type *, 4> Tys;
4943 for (Value *ArgOperand : CI->arg_operands())
4944 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4945
4946 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4947 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4948 ID == Intrinsic::lifetime_start)) {
4949 scalarizeInstruction(&I);
4950 break;
4951 }
4952 // The flag shows whether we use Intrinsic or a usual Call for vectorized
4953 // version of the instruction.
4954 // Is it beneficial to perform intrinsic call compared to lib call?
4955 bool NeedToScalarize;
4956 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4957 bool UseVectorIntrinsic =
4958 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4959 if (!UseVectorIntrinsic && NeedToScalarize) {
4960 scalarizeInstruction(&I);
4961 break;
4962 }
4963
4964 VectorParts Entry(UF);
4965 for (unsigned Part = 0; Part < UF; ++Part) {
4966 SmallVector<Value *, 4> Args;
4967 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
4968 Value *Arg = CI->getArgOperand(i);
4969 // Some intrinsics have a scalar argument - don't replace it with a
4970 // vector.
4971 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
4972 const VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
4973 Arg = VectorArg[Part];
4974 }
4975 Args.push_back(Arg);
4976 }
4977
4978 Function *VectorF;
4979 if (UseVectorIntrinsic) {
4980 // Use vector version of the intrinsic.
4981 Type *TysForDecl[] = {CI->getType()};
4982 if (VF > 1)
4983 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
4984 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
4985 } else {
4986 // Use vector version of the library call.
4987 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
4988 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4988, __PRETTY_FUNCTION__))
;
4989 VectorF = M->getFunction(VFnName);
4990 if (!VectorF) {
4991 // Generate a declaration
4992 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
4993 VectorF =
4994 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
4995 VectorF->copyAttributesFrom(F);
4996 }
4997 }
4998 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4998, __PRETTY_FUNCTION__))
;
4999
5000 SmallVector<OperandBundleDef, 1> OpBundles;
5001 CI->getOperandBundlesAsDefs(OpBundles);
5002 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
5003
5004 if (isa<FPMathOperator>(V))
5005 V->copyFastMathFlags(CI);
5006
5007 Entry[Part] = V;
5008 }
5009
5010 VectorLoopValueMap.initVector(&I, Entry);
5011 addMetadata(Entry, &I);
5012 break;
5013 }
5014
5015 default:
5016 // All other instructions are unsupported. Scalarize them.
5017 scalarizeInstruction(&I);
5018 break;
5019 } // end of switch.
5020 } // end of for_each instr.
5021}
5022
5023void InnerLoopVectorizer::updateAnalysis() {
5024 // Forget the original basic block.
5025 PSE.getSE()->forgetLoop(OrigLoop);
5026
5027 // Update the dominator tree information.
5028 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5029, __PRETTY_FUNCTION__))
5029 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5029, __PRETTY_FUNCTION__))
;
5030
5031 // We don't predicate stores by this point, so the vector body should be a
5032 // single loop.
5033 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
5034
5035 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody);
5036 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
5037 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
5038 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
5039
5040 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (false
)
;
5041}
5042
5043/// \brief Check whether it is safe to if-convert this phi node.
5044///
5045/// Phi nodes with constant expressions that can trap are not safe to if
5046/// convert.
5047static bool canIfConvertPHINodes(BasicBlock *BB) {
5048 for (Instruction &I : *BB) {
5049 auto *Phi = dyn_cast<PHINode>(&I);
5050 if (!Phi)
5051 return true;
5052 for (Value *V : Phi->incoming_values())
5053 if (auto *C = dyn_cast<Constant>(V))
5054 if (C->canTrap())
5055 return false;
5056 }
5057 return true;
5058}
5059
5060bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
5061 if (!EnableIfConversion) {
5062 ORE->emit(createMissedAnalysis("IfConversionDisabled")
5063 << "if-conversion is disabled");
5064 return false;
5065 }
5066
5067 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5067, __PRETTY_FUNCTION__))
;
5068
5069 // A list of pointers that we can safely read and write to.
5070 SmallPtrSet<Value *, 8> SafePointes;
5071
5072 // Collect safe addresses.
5073 for (BasicBlock *BB : TheLoop->blocks()) {
5074 if (blockNeedsPredication(BB))
5075 continue;
5076
5077 for (Instruction &I : *BB)
5078 if (auto *Ptr = getPointerOperand(&I))
5079 SafePointes.insert(Ptr);
5080 }
5081
5082 // Collect the blocks that need predication.
5083 BasicBlock *Header = TheLoop->getHeader();
5084 for (BasicBlock *BB : TheLoop->blocks()) {
5085 // We don't support switch statements inside loops.
5086 if (!isa<BranchInst>(BB->getTerminator())) {
5087 ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5088 << "loop contains a switch statement");
5089 return false;
5090 }
5091
5092 // We must be able to predicate all blocks that need to be predicated.
5093 if (blockNeedsPredication(BB)) {
5094 if (!blockCanBePredicated(BB, SafePointes)) {
5095 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5096 << "control flow cannot be substituted for a select");
5097 return false;
5098 }
5099 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5100 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5101 << "control flow cannot be substituted for a select");
5102 return false;
5103 }
5104 }
5105
5106 // We can if-convert this loop.
5107 return true;
5108}
5109
5110bool LoopVectorizationLegality::canVectorize() {
5111 // We must have a loop in canonical form. Loops with indirectbr in them cannot
5112 // be canonicalized.
5113 if (!TheLoop->getLoopPreheader()) {
5114 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5115 << "loop control flow is not understood by vectorizer");
5116 return false;
5117 }
5118
5119 // FIXME: The code is currently dead, since the loop gets sent to
5120 // LoopVectorizationLegality is already an innermost loop.
5121 //
5122 // We can only vectorize innermost loops.
5123 if (!TheLoop->empty()) {
5124 ORE->emit(createMissedAnalysis("NotInnermostLoop")
5125 << "loop is not the innermost loop");
5126 return false;
5127 }
5128
5129 // We must have a single backedge.
5130 if (TheLoop->getNumBackEdges() != 1) {
5131 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5132 << "loop control flow is not understood by vectorizer");
5133 return false;
5134 }
5135
5136 // We must have a single exiting block.
5137 if (!TheLoop->getExitingBlock()) {
5138 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5139 << "loop control flow is not understood by vectorizer");
5140 return false;
5141 }
5142
5143 // We only handle bottom-tested loops, i.e. loop in which the condition is
5144 // checked at the end of each iteration. With that we can assume that all
5145 // instructions in the loop are executed the same number of times.
5146 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5147 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5148 << "loop control flow is not understood by vectorizer");
5149 return false;
5150 }
5151
5152 // We need to have a loop header.
5153 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)
5154 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop: " <<
TheLoop->getHeader()->getName() << '\n'; } } while
(false)
;
5155
5156 // Check if we can if-convert non-single-bb loops.
5157 unsigned NumBlocks = TheLoop->getNumBlocks();
5158 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5159 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)
;
5160 return false;
5161 }
5162
5163 // ScalarEvolution needs to be able to find the exit count.
5164 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
5165 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
5166 ORE->emit(createMissedAnalysis("CantComputeNumberOfIterations")
5167 << "could not determine number of loop iterations");
5168 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)
;
5169 return false;
5170 }
5171
5172 // Check if we can vectorize the instructions and CFG in this loop.
5173 if (!canVectorizeInstrs()) {
5174 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)
;
5175 return false;
5176 }
5177
5178 // Go over each instruction and look at memory deps.
5179 if (!canVectorizeMemory()) {
5180 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)
;
5181 return false;
5182 }
5183
5184 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)
5185 << (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)
5186 ? " (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)
5187 : "")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)
5188 << "!\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)
;
5189
5190 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5191
5192 // If an override option has been passed in for interleaved accesses, use it.
5193 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5194 UseInterleaved = EnableInterleavedMemAccesses;
5195
5196 // Analyze interleaved memory accesses.
5197 if (UseInterleaved)
5198 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5199
5200 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5201 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5202 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5203
5204 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5205 ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5206 << "Too many SCEV assumptions need to be made and checked "
5207 << "at runtime");
5208 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)
;
5209 return false;
5210 }
5211
5212 // Okay! We can vectorize. At this point we don't have any other mem analysis
5213 // which may limit our maximum vectorization factor, so just return true with
5214 // no restrictions.
5215 return true;
5216}
5217
5218static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
5219 if (Ty->isPointerTy())
5220 return DL.getIntPtrType(Ty);
5221
5222 // It is possible that char's or short's overflow when we ask for the loop's
5223 // trip count, work around this by changing the type size.
5224 if (Ty->getScalarSizeInBits() < 32)
5225 return Type::getInt32Ty(Ty->getContext());
5226
5227 return Ty;
5228}
5229
5230static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
5231 Ty0 = convertPointerToIntegerType(DL, Ty0);
5232 Ty1 = convertPointerToIntegerType(DL, Ty1);
5233 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
5234 return Ty0;
5235 return Ty1;
5236}
5237
5238/// \brief Check that the instruction has outside loop users and is not an
5239/// identified reduction variable.
5240static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
5241 SmallPtrSetImpl<Value *> &AllowedExit) {
5242 // Reduction and Induction instructions are allowed to have exit users. All
5243 // other instructions must not have external users.
5244 if (!AllowedExit.count(Inst))
5245 // Check that all of the users of the loop are inside the BB.
5246 for (User *U : Inst->users()) {
5247 Instruction *UI = cast<Instruction>(U);
5248 // This user may be a reduction exit value.
5249 if (!TheLoop->contains(UI)) {
5250 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)
;
5251 return true;
5252 }
5253 }
5254 return false;
5255}
5256
5257void LoopVectorizationLegality::addInductionPhi(
5258 PHINode *Phi, const InductionDescriptor &ID,
5259 SmallPtrSetImpl<Value *> &AllowedExit) {
5260 Inductions[Phi] = ID;
5261 Type *PhiTy = Phi->getType();
5262 const DataLayout &DL = Phi->getModule()->getDataLayout();
5263
5264 // Get the widest type.
5265 if (!PhiTy->isFloatingPointTy()) {
5266 if (!WidestIndTy)
5267 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
5268 else
5269 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
5270 }
5271
5272 // Int inductions are special because we only allow one IV.
5273 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
5274 ID.getConstIntStepValue() &&
5275 ID.getConstIntStepValue()->isOne() &&
5276 isa<Constant>(ID.getStartValue()) &&
5277 cast<Constant>(ID.getStartValue())->isNullValue()) {
5278
5279 // Use the phi node with the widest type as induction. Use the last
5280 // one if there are multiple (no good reason for doing this other
5281 // than it is expedient). We've checked that it begins at zero and
5282 // steps by one, so this is a canonical induction variable.
5283 if (!PrimaryInduction || PhiTy == WidestIndTy)
5284 PrimaryInduction = Phi;
5285 }
5286
5287 // Both the PHI node itself, and the "post-increment" value feeding
5288 // back into the PHI node may have external users.
5289 AllowedExit.insert(Phi);
5290 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
5291
5292 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)
;
5293 return;
5294}
5295
5296bool LoopVectorizationLegality::canVectorizeInstrs() {
5297 BasicBlock *Header = TheLoop->getHeader();
5298
5299 // Look for the attribute signaling the absence of NaNs.
5300 Function &F = *Header->getParent();
5301 HasFunNoNaNAttr =
5302 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
5303
5304 // For each block in the loop.
5305 for (BasicBlock *BB : TheLoop->blocks()) {
5306 // Scan the instructions in the block and look for hazards.
5307 for (Instruction &I : *BB) {
5308 if (auto *Phi = dyn_cast<PHINode>(&I)) {
5309 Type *PhiTy = Phi->getType();
5310 // Check that this PHI type is allowed.
5311 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
5312 !PhiTy->isPointerTy()) {
5313 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5314 << "loop control flow is not understood by vectorizer");
5315 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)
;
5316 return false;
5317 }
5318
5319 // If this PHINode is not in the header block, then we know that we
5320 // can convert it to select during if-conversion. No need to check if
5321 // the PHIs in this block are induction or reduction variables.
5322 if (BB != Header) {
5323 // Check that this instruction has no outside users or is an
5324 // identified reduction value with an outside user.
5325 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
5326 continue;
5327 ORE->emit(createMissedAnalysis("NeitherInductionNorReduction", Phi)
5328 << "value could not be identified as "
5329 "an induction or reduction variable");
5330 return false;
5331 }
5332
5333 // We only allow if-converted PHIs with exactly two incoming values.
5334 if (Phi->getNumIncomingValues() != 2) {
5335 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5336 << "control flow not understood by vectorizer");
5337 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)
;
5338 return false;
5339 }
5340
5341 RecurrenceDescriptor RedDes;
5342 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
5343 if (RedDes.hasUnsafeAlgebra())
5344 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
5345 AllowedExit.insert(RedDes.getLoopExitInstr());
5346 Reductions[Phi] = RedDes;
5347 continue;
5348 }
5349
5350 InductionDescriptor ID;
5351 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
5352 addInductionPhi(Phi, ID, AllowedExit);
5353 if (ID.hasUnsafeAlgebra() && !HasFunNoNaNAttr)
5354 Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
5355 continue;
5356 }
5357
5358 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
5359 FirstOrderRecurrences.insert(Phi);
5360 continue;
5361 }
5362
5363 // As a last resort, coerce the PHI to a AddRec expression
5364 // and re-try classifying it a an induction PHI.
5365 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
5366 addInductionPhi(Phi, ID, AllowedExit);
5367 continue;
5368 }
5369
5370 ORE->emit(createMissedAnalysis("NonReductionValueUsedOutsideLoop", Phi)
5371 << "value that could not be identified as "
5372 "reduction is used outside the loop");
5373 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)
;
5374 return false;
5375 } // end of PHI handling
5376
5377 // We handle calls that:
5378 // * Are debug info intrinsics.
5379 // * Have a mapping to an IR intrinsic.
5380 // * Have a vector version available.
5381 auto *CI = dyn_cast<CallInst>(&I);
5382 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
5383 !isa<DbgInfoIntrinsic>(CI) &&
5384 !(CI->getCalledFunction() && TLI &&
5385 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
5386 ORE->emit(createMissedAnalysis("CantVectorizeCall", CI)
5387 << "call instruction cannot be vectorized");
5388 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)
;
5389 return false;
5390 }
5391
5392 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
5393 // second argument is the same (i.e. loop invariant)
5394 if (CI && hasVectorInstrinsicScalarOpd(
5395 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
5396 auto *SE = PSE.getSE();
5397 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
5398 ORE->emit(createMissedAnalysis("CantVectorizeIntrinsic", CI)
5399 << "intrinsic instruction cannot be vectorized");
5400 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)
;
5401 return false;
5402 }
5403 }
5404
5405 // Check that the instruction return type is vectorizable.
5406 // Also, we can't vectorize extractelement instructions.
5407 if ((!VectorType::isValidElementType(I.getType()) &&
5408 !I.getType()->isVoidTy()) ||
5409 isa<ExtractElementInst>(I)) {
5410 ORE->emit(createMissedAnalysis("CantVectorizeInstructionReturnType", &I)
5411 << "instruction return type cannot be vectorized");
5412 DEBUG(dbgs() << "LV: Found unvectorizable type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found unvectorizable type.\n"
; } } while (false)
;
5413 return false;
5414 }
5415
5416 // Check that the stored type is vectorizable.
5417 if (auto *ST = dyn_cast<StoreInst>(&I)) {
5418 Type *T = ST->getValueOperand()->getType();
5419 if (!VectorType::isValidElementType(T)) {
5420 ORE->emit(createMissedAnalysis("CantVectorizeStore", ST)
5421 << "store instruction cannot be vectorized");
5422 return false;
5423 }
5424
5425 // FP instructions can allow unsafe algebra, thus vectorizable by
5426 // non-IEEE-754 compliant SIMD units.
5427 // This applies to floating-point math operations and calls, not memory
5428 // operations, shuffles, or casts, as they don't change precision or
5429 // semantics.
5430 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
5431 !I.hasUnsafeAlgebra()) {
5432 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)
;
5433 Hints->setPotentiallyUnsafe();
5434 }
5435
5436 // Reduction instructions are allowed to have exit users.
5437 // All other instructions must not have external users.
5438 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
5439 ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)
5440 << "value cannot be used outside the loop");
5441 return false;
5442 }
5443
5444 } // next instr.
5445 }
5446
5447 if (!PrimaryInduction) {
5448 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)
;
5449 if (Inductions.empty()) {
5450 ORE->emit(createMissedAnalysis("NoInductionVariable")
5451 << "loop induction variable could not be identified");
5452 return false;
5453 }
5454 }
5455
5456 // Now we know the widest induction type, check if our found induction
5457 // is the same size. If it's not, unset it here and InnerLoopVectorizer
5458 // will create another.
5459 if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType())
5460 PrimaryInduction = nullptr;
5461
5462 return true;
5463}
5464
5465void LoopVectorizationCostModel::collectLoopScalars(unsigned VF) {
5466
5467 // We should not collect Scalars more than once per VF. Right now,
5468 // this function is called from collectUniformsAndScalars(), which
5469 // already does this check. Collecting Scalars for VF=1 does not make any
5470 // sense.
5471
5472 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5473, __PRETTY_FUNCTION__))
5473 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5473, __PRETTY_FUNCTION__))
;
5474
5475 // If an instruction is uniform after vectorization, it will remain scalar.
5476 Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end());
5477
5478 // Collect the getelementptr instructions that will not be vectorized. A
5479 // getelementptr instruction is only vectorized if it is used for a legal
5480 // gather or scatter operation.
5481 for (auto *BB : TheLoop->blocks())
5482 for (auto &I : *BB) {
5483 if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
5484 Scalars[VF].insert(GEP);
5485 continue;
5486 }
5487 auto *Ptr = getPointerOperand(&I);
5488 if (!Ptr)
5489 continue;
5490 auto *GEP = getGEPInstruction(Ptr);
5491 if (GEP && getWideningDecision(&I, VF) == CM_GatherScatter)
5492 Scalars[VF].erase(GEP);
5493 }
5494
5495 // An induction variable will remain scalar if all users of the induction
5496 // variable and induction variable update remain scalar.
5497 auto *Latch = TheLoop->getLoopLatch();
5498 for (auto &Induction : *Legal->getInductionVars()) {
5499 auto *Ind = Induction.first;
5500 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5501
5502 // Determine if all users of the induction variable are scalar after
5503 // vectorization.
5504 auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
5505 auto *I = cast<Instruction>(U);
5506 return I == IndUpdate || !TheLoop->contains(I) || Scalars[VF].count(I);
5507 });
5508 if (!ScalarInd)
5509 continue;
5510
5511 // Determine if all users of the induction variable update instruction are
5512 // scalar after vectorization.
5513 auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5514 auto *I = cast<Instruction>(U);
5515 return I == Ind || !TheLoop->contains(I) || Scalars[VF].count(I);
5516 });
5517 if (!ScalarIndUpdate)
5518 continue;
5519
5520 // The induction variable and its update instruction will remain scalar.
5521 Scalars[VF].insert(Ind);
5522 Scalars[VF].insert(IndUpdate);
5523 }
5524}
5525
5526bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
5527 if (!blockNeedsPredication(I->getParent()))
5528 return false;
5529 switch(I->getOpcode()) {
5530 default:
5531 break;
5532 case Instruction::Store:
5533 return !isMaskRequired(I);
5534 case Instruction::UDiv:
5535 case Instruction::SDiv:
5536 case Instruction::SRem:
5537 case Instruction::URem:
5538 return mayDivideByZero(*I);
5539 }
5540 return false;
5541}
5542
5543bool LoopVectorizationLegality::memoryInstructionCanBeWidened(Instruction *I,
5544 unsigned VF) {
5545 // Get and ensure we have a valid memory instruction.
5546 LoadInst *LI = dyn_cast<LoadInst>(I);
5547 StoreInst *SI = dyn_cast<StoreInst>(I);
5548 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5548, __PRETTY_FUNCTION__))
;
5549
5550 auto *Ptr = getPointerOperand(I);
5551
5552 // In order to be widened, the pointer should be consecutive, first of all.
5553 if (!isConsecutivePtr(Ptr))
5554 return false;
5555
5556 // If the instruction is a store located in a predicated block, it will be
5557 // scalarized.
5558 if (isScalarWithPredication(I))
5559 return false;
5560
5561 // If the instruction's allocated size doesn't equal it's type size, it
5562 // requires padding and will be scalarized.
5563 auto &DL = I->getModule()->getDataLayout();
5564 auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
5565 if (hasIrregularType(ScalarTy, DL, VF))
5566 return false;
5567
5568 return true;
5569}
5570
5571void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) {
5572
5573 // We should not collect Uniforms more than once per VF. Right now,
5574 // this function is called from collectUniformsAndScalars(), which
5575 // already does this check. Collecting Uniforms for VF=1 does not make any
5576 // sense.
5577
5578 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5579, __PRETTY_FUNCTION__))
5579 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5579, __PRETTY_FUNCTION__))
;
5580
5581 // Visit the list of Uniforms. If we'll not find any uniform value, we'll
5582 // not analyze again. Uniforms.count(VF) will return 1.
5583 Uniforms[VF].clear();
5584
5585 // We now know that the loop is vectorizable!
5586 // Collect instructions inside the loop that will remain uniform after
5587 // vectorization.
5588
5589 // Global values, params and instructions outside of current loop are out of
5590 // scope.
5591 auto isOutOfScope = [&](Value *V) -> bool {
5592 Instruction *I = dyn_cast<Instruction>(V);
5593 return (!I || !TheLoop->contains(I));
5594 };
5595
5596 SetVector<Instruction *> Worklist;
5597 BasicBlock *Latch = TheLoop->getLoopLatch();
5598
5599 // Start with the conditional branch. If the branch condition is an
5600 // instruction contained in the loop that is only used by the branch, it is
5601 // uniform.
5602 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
5603 if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) {
5604 Worklist.insert(Cmp);
5605 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)
;
5606 }
5607
5608 // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
5609 // are pointers that are treated like consecutive pointers during
5610 // vectorization. The pointer operands of interleaved accesses are an
5611 // example.
5612 SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
5613
5614 // Holds pointer operands of instructions that are possibly non-uniform.
5615 SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
5616
5617 auto isUniformDecision = [&](Instruction *I, unsigned VF) {
5618 InstWidening WideningDecision = getWideningDecision(I, VF);
5619 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5620, __PRETTY_FUNCTION__))
5620 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5620, __PRETTY_FUNCTION__))
;
5621
5622 return (WideningDecision == CM_Widen ||
5623 WideningDecision == CM_Interleave);
5624 };
5625 // Iterate over the instructions in the loop, and collect all
5626 // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
5627 // that a consecutive-like pointer operand will be scalarized, we collect it
5628 // in PossibleNonUniformPtrs instead. We use two sets here because a single
5629 // getelementptr instruction can be used by both vectorized and scalarized
5630 // memory instructions. For example, if a loop loads and stores from the same
5631 // location, but the store is conditional, the store will be scalarized, and
5632 // the getelementptr won't remain uniform.
5633 for (auto *BB : TheLoop->blocks())
5634 for (auto &I : *BB) {
5635
5636 // If there's no pointer operand, there's nothing to do.
5637 auto *Ptr = dyn_cast_or_null<Instruction>(getPointerOperand(&I));
5638 if (!Ptr)
5639 continue;
5640
5641 // True if all users of Ptr are memory accesses that have Ptr as their
5642 // pointer operand.
5643 auto UsersAreMemAccesses = all_of(Ptr->users(), [&](User *U) -> bool {
5644 return getPointerOperand(U) == Ptr;
5645 });
5646
5647 // Ensure the memory instruction will not be scalarized or used by
5648 // gather/scatter, making its pointer operand non-uniform. If the pointer
5649 // operand is used by any instruction other than a memory access, we
5650 // conservatively assume the pointer operand may be non-uniform.
5651 if (!UsersAreMemAccesses || !isUniformDecision(&I, VF))
5652 PossibleNonUniformPtrs.insert(Ptr);
5653
5654 // If the memory instruction will be vectorized and its pointer operand
5655 // is consecutive-like, or interleaving - the pointer operand should
5656 // remain uniform.
5657 else
5658 ConsecutiveLikePtrs.insert(Ptr);
5659 }
5660
5661 // Add to the Worklist all consecutive and consecutive-like pointers that
5662 // aren't also identified as possibly non-uniform.
5663 for (auto *V : ConsecutiveLikePtrs)
5664 if (!PossibleNonUniformPtrs.count(V)) {
5665 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)
;
5666 Worklist.insert(V);
5667 }
5668
5669 // Expand Worklist in topological order: whenever a new instruction
5670 // is added , its users should be either already inside Worklist, or
5671 // out of scope. It ensures a uniform instruction will only be used
5672 // by uniform instructions or out of scope instructions.
5673 unsigned idx = 0;
5674 while (idx != Worklist.size()) {
5675 Instruction *I = Worklist[idx++];
5676
5677 for (auto OV : I->operand_values()) {
5678 if (isOutOfScope(OV))
5679 continue;
5680 auto *OI = cast<Instruction>(OV);
5681 if (all_of(OI->users(), [&](User *U) -> bool {
5682 return isOutOfScope(U) || Worklist.count(cast<Instruction>(U));
5683 })) {
5684 Worklist.insert(OI);
5685 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)
;
5686 }
5687 }
5688 }
5689
5690 // Returns true if Ptr is the pointer operand of a memory access instruction
5691 // I, and I is known to not require scalarization.
5692 auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
5693 return getPointerOperand(I) == Ptr && isUniformDecision(I, VF);
5694 };
5695
5696 // For an instruction to be added into Worklist above, all its users inside
5697 // the loop should also be in Worklist. However, this condition cannot be
5698 // true for phi nodes that form a cyclic dependence. We must process phi
5699 // nodes separately. An induction variable will remain uniform if all users
5700 // of the induction variable and induction variable update remain uniform.
5701 // The code below handles both pointer and non-pointer induction variables.
5702 for (auto &Induction : *Legal->getInductionVars()) {
5703 auto *Ind = Induction.first;
5704 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5705
5706 // Determine if all users of the induction variable are uniform after
5707 // vectorization.
5708 auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
5709 auto *I = cast<Instruction>(U);
5710 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
5711 isVectorizedMemAccessUse(I, Ind);
5712 });
5713 if (!UniformInd)
5714 continue;
5715
5716 // Determine if all users of the induction variable update instruction are
5717 // uniform after vectorization.
5718 auto UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5719 auto *I = cast<Instruction>(U);
5720 return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||
5721 isVectorizedMemAccessUse(I, IndUpdate);
5722 });
5723 if (!UniformIndUpdate)
5724 continue;
5725
5726 // The induction variable and its update instruction will remain uniform.
5727 Worklist.insert(Ind);
5728 Worklist.insert(IndUpdate);
5729 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)
;
5730 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)
;
5731 }
5732
5733 Uniforms[VF].insert(Worklist.begin(), Worklist.end());
5734}
5735
5736bool LoopVectorizationLegality::canVectorizeMemory() {
5737 LAI = &(*GetLAA)(*TheLoop);
5738 InterleaveInfo.setLAI(LAI);
5739 const OptimizationRemarkAnalysis *LAR = LAI->getReport();
5740 if (LAR) {
5741 OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
5742 "loop not vectorized: ", *LAR);
5743 ORE->emit(VR);
5744 }
5745 if (!LAI->canVectorizeMemory())
5746 return false;
5747
5748 if (LAI->hasStoreToLoopInvariantAddress()) {
5749 ORE->emit(createMissedAnalysis("CantVectorizeStoreToLoopInvariantAddress")
5750 << "write to a loop invariant address could not be vectorized");
5751 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)
;
5752 return false;
5753 }
5754
5755 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
5756 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
5757
5758 return true;
5759}
5760
5761bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5762 Value *In0 = const_cast<Value *>(V);
5763 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5764 if (!PN)
5765 return false;
5766
5767 return Inductions.count(PN);
5768}
5769
5770bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
5771 return FirstOrderRecurrences.count(Phi);
5772}
5773
5774bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5775 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
5776}
5777
5778bool LoopVectorizationLegality::blockCanBePredicated(
5779 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
5780 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
5781
5782 for (Instruction &I : *BB) {
5783 // Check that we don't have a constant expression that can trap as operand.
5784 for (Value *Operand : I.operands()) {
5785 if (auto *C = dyn_cast<Constant>(Operand))
5786 if (C->canTrap())
5787 return false;
5788 }
5789 // We might be able to hoist the load.
5790 if (I.mayReadFromMemory()) {
5791 auto *LI = dyn_cast<LoadInst>(&I);
5792 if (!LI)
5793 return false;
5794 if (!SafePtrs.count(LI->getPointerOperand())) {
5795 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
5796 isLegalMaskedGather(LI->getType())) {
5797 MaskedOp.insert(LI);
5798 continue;
5799 }
5800 // !llvm.mem.parallel_loop_access implies if-conversion safety.
5801 if (IsAnnotatedParallel)
5802 continue;
5803 return false;
5804 }
5805 }
5806
5807 if (I.mayWriteToMemory()) {
5808 auto *SI = dyn_cast<StoreInst>(&I);
5809 // We only support predication of stores in basic blocks with one
5810 // predecessor.
5811 if (!SI)
5812 return false;
5813
5814 // Build a masked store if it is legal for the target.
5815 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
5816 SI->getPointerOperand()) ||
5817 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
5818 MaskedOp.insert(SI);
5819 continue;
5820 }
5821
5822 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5823 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5824
5825 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5826 !isSinglePredecessor)
5827 return false;
5828 }
5829 if (I.mayThrow())
5830 return false;
5831 }
5832
5833 return true;
5834}
5835
5836void InterleavedAccessInfo::collectConstStrideAccesses(
5837 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
5838 const ValueToValueMap &Strides) {
5839
5840 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
5841
5842 // Since it's desired that the load/store instructions be maintained in
5843 // "program order" for the interleaved access analysis, we have to visit the
5844 // blocks in the loop in reverse postorder (i.e., in a topological order).
5845 // Such an ordering will ensure that any load/store that may be executed
5846 // before a second load/store will precede the second load/store in
5847 // AccessStrideInfo.
5848 LoopBlocksDFS DFS(TheLoop);
5849 DFS.perform(LI);
5850 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
5851 for (auto &I : *BB) {
5852 auto *LI = dyn_cast<LoadInst>(&I);
5853 auto *SI = dyn_cast<StoreInst>(&I);
5854 if (!LI && !SI)
5855 continue;
5856
5857 Value *Ptr = getPointerOperand(&I);
5858 // We don't check wrapping here because we don't know yet if Ptr will be
5859 // part of a full group or a group with gaps. Checking wrapping for all
5860 // pointers (even those that end up in groups with no gaps) will be overly
5861 // conservative. For full groups, wrapping should be ok since if we would
5862 // wrap around the address space we would do a memory access at nullptr
5863 // even without the transformation. The wrapping checks are therefore
5864 // deferred until after we've formed the interleaved groups.
5865 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
5866 /*Assume=*/true, /*ShouldCheckWrap=*/false);
5867
5868 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
5869 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5870 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
5871
5872 // An alignment of 0 means target ABI alignment.
5873 unsigned Align = getMemInstAlignment(&I);
5874 if (!Align)
5875 Align = DL.getABITypeAlignment(PtrTy->getElementType());
5876
5877 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
5878 }
5879}
5880
5881// Analyze interleaved accesses and collect them into interleaved load and
5882// store groups.
5883//
5884// When generating code for an interleaved load group, we effectively hoist all
5885// loads in the group to the location of the first load in program order. When
5886// generating code for an interleaved store group, we sink all stores to the
5887// location of the last store. This code motion can change the order of load
5888// and store instructions and may break dependences.
5889//
5890// The code generation strategy mentioned above ensures that we won't violate
5891// any write-after-read (WAR) dependences.
5892//
5893// E.g., for the WAR dependence: a = A[i]; // (1)
5894// A[i] = b; // (2)
5895//
5896// The store group of (2) is always inserted at or below (2), and the load
5897// group of (1) is always inserted at or above (1). Thus, the instructions will
5898// never be reordered. All other dependences are checked to ensure the
5899// correctness of the instruction reordering.
5900//
5901// The algorithm visits all memory accesses in the loop in bottom-up program
5902// order. Program order is established by traversing the blocks in the loop in
5903// reverse postorder when collecting the accesses.
5904//
5905// We visit the memory accesses in bottom-up order because it can simplify the
5906// construction of store groups in the presence of write-after-write (WAW)
5907// dependences.
5908//
5909// E.g., for the WAW dependence: A[i] = a; // (1)
5910// A[i] = b; // (2)
5911// A[i + 1] = c; // (3)
5912//
5913// We will first create a store group with (3) and (2). (1) can't be added to
5914// this group because it and (2) are dependent. However, (1) can be grouped
5915// with other accesses that may precede it in program order. Note that a
5916// bottom-up order does not imply that WAW dependences should not be checked.
5917void InterleavedAccessInfo::analyzeInterleaving(
5918 const ValueToValueMap &Strides) {
5919 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Analyzing interleaved accesses...\n"
; } } while (false)
;
5920
5921 // Holds all accesses with a constant stride.
5922 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
5923 collectConstStrideAccesses(AccessStrideInfo, Strides);
5924
5925 if (AccessStrideInfo.empty())
1
Assuming the condition is false
2
Taking false branch
5926 return;
5927
5928 // Collect the dependences in the loop.
5929 collectDependences();
5930
5931 // Holds all interleaved store groups temporarily.
5932 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
5933 // Holds all interleaved load groups temporarily.
5934 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
5935
5936 // Search in bottom-up program order for pairs of accesses (A and B) that can
5937 // form interleaved load or store groups. In the algorithm below, access A
5938 // precedes access B in program order. We initialize a group for B in the
5939 // outer loop of the algorithm, and then in the inner loop, we attempt to
5940 // insert each A into B's group if:
5941 //
5942 // 1. A and B have the same stride,
5943 // 2. A and B have the same memory object size, and
5944 // 3. A belongs in B's group according to its distance from B.
5945 //
5946 // Special care is taken to ensure group formation will not break any
5947 // dependences.
5948 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
3
Loop condition is true. Entering loop body
5949 BI != E; ++BI) {
5950 Instruction *B = BI->first;
5951 StrideDescriptor DesB = BI->second;
5952
5953 // Initialize a group for B if it has an allowable stride. Even if we don't
5954 // create a group for B, we continue with the bottom-up algorithm to ensure
5955 // we don't break any of B's dependences.
5956 InterleaveGroup *Group = nullptr;
4
'Group' initialized to a null pointer value
5957 if (isStrided(DesB.Stride)) {
5
Taking false branch
5958 Group = getInterleaveGroup(B);
5959 if (!Group) {
5960 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)
;
5961 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
5962 }
5963 if (B->mayWriteToMemory())
5964 StoreGroups.insert(Group);
5965 else
5966 LoadGroups.insert(Group);
5967 }
5968
5969 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
5970 Instruction *A = AI->first;
5971 StrideDescriptor DesA = AI->second;
5972
5973 // Our code motion strategy implies that we can't have dependences
5974 // between accesses in an interleaved group and other accesses located
5975 // between the first and last member of the group. Note that this also
5976 // means that a group can't have more than one member at a given offset.
5977 // The accesses in a group can have dependences with other accesses, but
5978 // we must ensure we don't extend the boundaries of the group such that
5979 // we encompass those dependent accesses.
5980 //
5981 // For example, assume we have the sequence of accesses shown below in a
5982 // stride-2 loop:
5983 //
5984 // (1, 2) is a group | A[i] = a; // (1)
5985 // | A[i-1] = b; // (2) |
5986 // A[i-3] = c; // (3)
5987 // A[i] = d; // (4) | (2, 4) is not a group
5988 //
5989 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
5990 // but not with (4). If we did, the dependent access (3) would be within
5991 // the boundaries of the (2, 4) group.
5992 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
7
Taking false branch
10
Taking false branch
13
Taking false branch
16
Taking false branch
5993
5994 // If a dependence exists and A is already in a group, we know that A
5995 // must be a store since A precedes B and WAR dependences are allowed.
5996 // Thus, A would be sunk below B. We release A's group to prevent this
5997 // illegal code motion. A will then be free to form another group with
5998 // instructions that precede it.
5999 if (isInterleaved(A)) {
6000 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
6001 StoreGroups.remove(StoreGroup);
6002 releaseGroup(StoreGroup);
6003 }
6004
6005 // If a dependence exists and A is not already in a group (or it was
6006 // and we just released it), B might be hoisted above A (if B is a
6007 // load) or another store might be sunk below A (if B is a store). In
6008 // either case, we can't add additional instructions to B's group. B
6009 // will only form a group with instructions that it precedes.
6010 break;
6011 }
6012
6013 // At this point, we've checked for illegal code motion. If either A or B
6014 // isn't strided, there's nothing left to do.
6015 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
17
Taking false branch
6016 continue;
8
Execution continues on line 5969
11
Execution continues on line 5969
14
Execution continues on line 5969
6017
6018 // Ignore A if it's already in a group or isn't the same kind of memory
6019 // operation as B.
6020 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
18
Assuming the condition is false
19
Taking false branch
6021 continue;
6022
6023 // Check rules 1 and 2. Ignore A if its stride or size is different from
6024 // that of B.
6025 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
20
Taking false branch
6026 continue;
6027
6028 // Calculate the distance from A to B.
6029 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
6030 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
6031 if (!DistToB)
21
Assuming 'DistToB' is non-null
22
Taking false branch
6032 continue;
6033 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
6034
6035 // Check rule 3. Ignore A if its distance to B is not a multiple of the
6036 // size.
6037 if (DistanceToB % static_cast<int64_t>(DesB.Size))
23
Taking false branch
6038 continue;
6039
6040 // Ignore A if either A or B is in a predicated block. Although we
6041 // currently prevent group formation for predicated accesses, we may be
6042 // able to relax this limitation in the future once we handle more
6043 // complicated blocks.
6044 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
24
Assuming the condition is false
25
Assuming the condition is false
26
Taking false branch
6045 continue;
6046
6047 // The index of A is the index of B plus A's distance to B in multiples
6048 // of the size.
6049 int IndexA =
6050 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
27
Called C++ object pointer is null
6051
6052 // Try to insert A into B's group.
6053 if (Group->insertMember(A, IndexA, DesA.Align)) {
6054 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)
6055 << " 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)
;
6056 InterleaveGroupMap[A] = Group;
6057
6058 // Set the first load in program order as the insert position.
6059 if (A->mayReadFromMemory())
6060 Group->setInsertPos(A);
6061 }
6062 } // Iteration over A accesses.
6063 } // Iteration over B accesses.
6064
6065 // Remove interleaved store groups with gaps.
6066 for (InterleaveGroup *Group : StoreGroups)
6067 if (Group->getNumMembers() != Group->getFactor())
6068 releaseGroup(Group);
6069
6070 // Remove interleaved groups with gaps (currently only loads) whose memory
6071 // accesses may wrap around. We have to revisit the getPtrStride analysis,
6072 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
6073 // not check wrapping (see documentation there).
6074 // FORNOW we use Assume=false;
6075 // TODO: Change to Assume=true but making sure we don't exceed the threshold
6076 // of runtime SCEV assumptions checks (thereby potentially failing to
6077 // vectorize altogether).
6078 // Additional optional optimizations:
6079 // TODO: If we are peeling the loop and we know that the first pointer doesn't
6080 // wrap then we can deduce that all pointers in the group don't wrap.
6081 // This means that we can forcefully peel the loop in order to only have to
6082 // check the first pointer for no-wrap. When we'll change to use Assume=true
6083 // we'll only need at most one runtime check per interleaved group.
6084 //
6085 for (InterleaveGroup *Group : LoadGroups) {
6086
6087 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
6088 // load would wrap around the address space we would do a memory access at
6089 // nullptr even without the transformation.
6090 if (Group->getNumMembers() == Group->getFactor())
6091 continue;
6092
6093 // Case 2: If first and last members of the group don't wrap this implies
6094 // that all the pointers in the group don't wrap.
6095 // So we check only group member 0 (which is always guaranteed to exist),
6096 // and group member Factor - 1; If the latter doesn't exist we rely on
6097 // peeling (if it is a non-reveresed accsess -- see Case 3).
6098 Value *FirstMemberPtr = getPointerOperand(Group->getMember(0));
6099 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
6100 /*ShouldCheckWrap=*/true)) {
6101 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)
6102 "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)
;
6103 releaseGroup(Group);
6104 continue;
6105 }
6106 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
6107 if (LastMember) {
6108 Value *LastMemberPtr = getPointerOperand(LastMember);
6109 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
6110 /*ShouldCheckWrap=*/true)) {
6111 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)
6112 "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)
;
6113 releaseGroup(Group);
6114 }
6115 } else {
6116 // Case 3: A non-reversed interleaved load group with gaps: We need
6117 // to execute at least one scalar epilogue iteration. This will ensure
6118 // we don't speculatively access memory out-of-bounds. We only need
6119 // to look for a member at index factor - 1, since every group must have
6120 // a member at index zero.
6121 if (Group->isReverse()) {
6122 releaseGroup(Group);
6123 continue;
6124 }
6125 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)
;
6126 RequiresScalarEpilogue = true;
6127 }
6128 }
6129}
6130
6131LoopVectorizationCostModel::VectorizationFactor
6132LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
6133 // Width 1 means no vectorize
6134 VectorizationFactor Factor = {1U, 0U};
6135 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
6136 ORE->emit(createMissedAnalysis("CantVersionLoopWithOptForSize")
6137 << "runtime pointer checks needed. Enable vectorization of this "
6138 "loop with '#pragma clang loop vectorize(enable)' when "
6139 "compiling with -Os/-Oz");
6140 DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n"
; } } while (false)
6141 << "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)
;
6142 return Factor;
6143 }
6144
6145 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
6146 ORE->emit(createMissedAnalysis("ConditionalStore")
6147 << "store that is conditionally executed prevents vectorization");
6148 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)
;
6149 return Factor;
6150 }
6151
6152 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
6153 unsigned SmallestType, WidestType;
6154 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
6155 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
6156 unsigned MaxSafeDepDist = -1U;
6157
6158 // Get the maximum safe dependence distance in bits computed by LAA. If the
6159 // loop contains any interleaved accesses, we divide the dependence distance
6160 // by the maximum interleave factor of all interleaved groups. Note that
6161 // although the division ensures correctness, this is a fairly conservative
6162 // computation because the maximum distance computed by LAA may not involve
6163 // any of the interleaved accesses.
6164 if (Legal->getMaxSafeDepDistBytes() != -1U)
6165 MaxSafeDepDist =
6166 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
6167
6168 WidestRegister =
6169 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
6170 unsigned MaxVectorSize = WidestRegister / WidestType;
6171
6172 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)
6173 << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: "
<< SmallestType << " / " << WidestType <<
" bits.\n"; } } while (false)
;
6174 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
)
6175 << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest register is: "
<< WidestRegister << " bits.\n"; } } while (false
)
;
6176
6177 if (MaxVectorSize == 0) {
6178 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)
;
6179 MaxVectorSize = 1;
6180 }
6181
6182 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6183, __PRETTY_FUNCTION__))
6183 " 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6183, __PRETTY_FUNCTION__))
;
6184
6185 unsigned VF = MaxVectorSize;
6186 if (MaximizeBandwidth && !OptForSize) {
6187 // Collect all viable vectorization factors.
6188 SmallVector<unsigned, 8> VFs;
6189 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
6190 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
6191 VFs.push_back(VS);
6192
6193 // For each VF calculate its register usage.
6194 auto RUs = calculateRegisterUsage(VFs);
6195
6196 // Select the largest VF which doesn't require more registers than existing
6197 // ones.
6198 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
6199 for (int i = RUs.size() - 1; i >= 0; --i) {
6200 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
6201 VF = VFs[i];
6202 break;
6203 }
6204 }
6205 }
6206
6207 // If we optimize the program for size, avoid creating the tail loop.
6208 if (OptForSize) {
6209 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6210 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)
;
6211
6212 // If we don't know the precise trip count, don't try to vectorize.
6213 if (TC < 2) {
6214 ORE->emit(
6215 createMissedAnalysis("UnknownLoopCountComplexCFG")
6216 << "unable to calculate the loop count due to complex control flow");
6217 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)
;
6218 return Factor;
6219 }
6220
6221 // Find the maximum SIMD width that can fit within the trip count.
6222 VF = TC % MaxVectorSize;
6223
6224 if (VF == 0)
6225 VF = MaxVectorSize;
6226 else {
6227 // If the trip count that we found modulo the vectorization factor is not
6228 // zero then we require a tail.
6229 ORE->emit(createMissedAnalysis("NoTailLoopWithOptForSize")
6230 << "cannot optimize for size and vectorize at the "
6231 "same time. Enable vectorization of this loop "
6232 "with '#pragma clang loop vectorize(enable)' "
6233 "when compiling with -Os/-Oz");
6234 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)
;
6235 return Factor;
6236 }
6237 }
6238
6239 int UserVF = Hints->getWidth();
6240 if (UserVF != 0) {
6241 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two")((isPowerOf2_32(UserVF) && "VF needs to be a power of two"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(UserVF) && \"VF needs to be a power of two\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6241, __PRETTY_FUNCTION__))
;
6242 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Using user VF " <<
UserVF << ".\n"; } } while (false)
;
6243
6244 Factor.Width = UserVF;
6245
6246 collectUniformsAndScalars(UserVF);
6247 collectInstsToScalarize(UserVF);
6248 return Factor;
6249 }
6250
6251 float Cost = expectedCost(1).first;
6252#ifndef NDEBUG
6253 const float ScalarCost = Cost;
6254#endif /* NDEBUG */
6255 unsigned Width = 1;
6256 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)
;
6257
6258 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
6259 // Ignore scalar width, because the user explicitly wants vectorization.
6260 if (ForceVectorization && VF > 1) {
6261 Width = 2;
6262 Cost = expectedCost(Width).first / (float)Width;
6263 }
6264
6265 for (unsigned i = 2; i <= VF; i *= 2) {
6266 // Notice that the vector loop needs to be executed less times, so
6267 // we need to divide the cost of the vector loops by the width of
6268 // the vector elements.
6269 VectorizationCostTy C = expectedCost(i);
6270 float VectorCost = C.first / (float)i;
6271 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)
6272 << " 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)
;
6273 if (!C.second && !ForceVectorization) {
6274 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)
6275 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)
6276 << " 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)
;
6277 continue;
6278 }
6279 if (VectorCost < Cost) {
6280 Cost = VectorCost;
6281 Width = i;
6282 }
6283 }
6284
6285 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)
6286 << "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)
6287 << "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)
;
6288 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Selecting VF: " <<
Width << ".\n"; } } while (false)
;
6289 Factor.Width = Width;
6290 Factor.Cost = Width * Cost;
6291 return Factor;
6292}
6293
6294std::pair<unsigned, unsigned>
6295LoopVectorizationCostModel::getSmallestAndWidestTypes() {
6296 unsigned MinWidth = -1U;
6297 unsigned MaxWidth = 8;
6298 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6299
6300 // For each block.
6301 for (BasicBlock *BB : TheLoop->blocks()) {
6302 // For each instruction in the loop.
6303 for (Instruction &I : *BB) {
6304 Type *T = I.getType();
6305
6306 // Skip ignored values.
6307 if (ValuesToIgnore.count(&I))
6308 continue;
6309
6310 // Only examine Loads, Stores and PHINodes.
6311 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
6312 continue;
6313
6314 // Examine PHI nodes that are reduction variables. Update the type to
6315 // account for the recurrence type.
6316 if (auto *PN = dyn_cast<PHINode>(&I)) {
6317 if (!Legal->isReductionVariable(PN))
6318 continue;
6319 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
6320 T = RdxDesc.getRecurrenceType();
6321 }
6322
6323 // Examine the stored values.
6324 if (auto *ST = dyn_cast<StoreInst>(&I))
6325 T = ST->getValueOperand()->getType();
6326
6327 // Ignore loaded pointer types and stored pointer types that are not
6328 // consecutive. However, we do want to take consecutive stores/loads of
6329 // pointer vectors into account.
6330 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I))
6331 continue;
6332
6333 MinWidth = std::min(MinWidth,
6334 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6335 MaxWidth = std::max(MaxWidth,
6336 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6337 }
6338 }
6339
6340 return {MinWidth, MaxWidth};
6341}
6342
6343unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
6344 unsigned VF,
6345 unsigned LoopCost) {
6346
6347 // -- The interleave heuristics --
6348 // We interleave the loop in order to expose ILP and reduce the loop overhead.
6349 // There are many micro-architectural considerations that we can't predict
6350 // at this level. For example, frontend pressure (on decode or fetch) due to
6351 // code size, or the number and capabilities of the execution ports.
6352 //
6353 // We use the following heuristics to select the interleave count:
6354 // 1. If the code has reductions, then we interleave to break the cross
6355 // iteration dependency.
6356 // 2. If the loop is really small, then we interleave to reduce the loop
6357 // overhead.
6358 // 3. We don't interleave if we think that we will spill registers to memory
6359 // due to the increased register pressure.
6360
6361 // When we optimize for size, we don't interleave.
6362 if (OptForSize)
6363 return 1;
6364
6365 // We used the distance for the interleave count.
6366 if (Legal->getMaxSafeDepDistBytes() != -1U)
6367 return 1;
6368
6369 // Do not interleave loops with a relatively small trip count.
6370 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6371 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
6372 return 1;
6373
6374 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
6375 DEBUG(dbgs() << "LV: The target has " << TargetNumRegistersdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (false
)
6376 << " registers\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (false
)
;
6377
6378 if (VF == 1) {
6379 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
6380 TargetNumRegisters = ForceTargetNumScalarRegs;
6381 } else {
6382 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
6383 TargetNumRegisters = ForceTargetNumVectorRegs;
6384 }
6385
6386 RegisterUsage R = calculateRegisterUsage({VF})[0];
6387 // We divide by these constants so assume that we have at least one
6388 // instruction that uses at least one register.
6389 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
6390 R.NumInstructions = std::max(R.NumInstructions, 1U);
6391
6392 // We calculate the interleave count using the following formula.
6393 // Subtract the number of loop invariants from the number of available
6394 // registers. These registers are used by all of the interleaved instances.
6395 // Next, divide the remaining registers by the number of registers that is
6396 // required by the loop, in order to estimate how many parallel instances
6397 // fit without causing spills. All of this is rounded down if necessary to be
6398 // a power of two. We want power of two interleave count to simplify any
6399 // addressing operations or alignment considerations.
6400 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
6401 R.MaxLocalUsers);
6402
6403 // Don't count the induction variable as interleaved.
6404 if (EnableIndVarRegisterHeur)
6405 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
6406 std::max(1U, (R.MaxLocalUsers - 1)));
6407
6408 // Clamp the interleave ranges to reasonable counts.
6409 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
6410
6411 // Check if the user has overridden the max.
6412 if (VF == 1) {
6413 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
6414 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
6415 } else {
6416 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
6417 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
6418 }
6419
6420 // If we did not calculate the cost for VF (because the user selected the VF)
6421 // then we calculate the cost of VF here.
6422 if (LoopCost == 0)
6423 LoopCost = expectedCost(VF).first;
6424
6425 // Clamp the calculated IC to be between the 1 and the max interleave count
6426 // that the target allows.
6427 if (IC > MaxInterleaveCount)
6428 IC = MaxInterleaveCount;
6429 else if (IC < 1)
6430 IC = 1;
6431
6432 // Interleave if we vectorized this loop and there is a reduction that could
6433 // benefit from interleaving.
6434 if (VF > 1 && Legal->getReductionVars()->size()) {
6435 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)
;
6436 return IC;
6437 }
6438
6439 // Note that if we've already vectorized the loop we will have done the
6440 // runtime check and so interleaving won't require further checks.
6441 bool InterleavingRequiresRuntimePointerCheck =
6442 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
6443
6444 // We want to interleave small loops in order to reduce the loop overhead and
6445 // potentially expose ILP opportunities.
6446 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)
;
6447 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
6448 // We assume that the cost overhead is 1 and we use the cost model
6449 // to estimate the cost of the loop and interleave until the cost of the
6450 // loop overhead is about 5% of the cost of the loop.
6451 unsigned SmallIC =
6452 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
6453
6454 // Interleave until store/load ports (estimated by max interleave count) are
6455 // saturated.
6456 unsigned NumStores = Legal->getNumStores();
6457 unsigned NumLoads = Legal->getNumLoads();
6458 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
6459 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
6460
6461 // If we have a scalar reduction (vector reductions are already dealt with
6462 // by this point), we can increase the critical path length if the loop
6463 // we're interleaving is inside another loop. Limit, by default to 2, so the
6464 // critical path only gets increased by one reduction operation.
6465 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
6466 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
6467 SmallIC = std::min(SmallIC, F);
6468 StoresIC = std::min(StoresIC, F);
6469 LoadsIC = std::min(LoadsIC, F);
6470 }
6471
6472 if (EnableLoadStoreRuntimeInterleave &&
6473 std::max(StoresIC, LoadsIC) > SmallIC) {
6474 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)
;
6475 return std::max(StoresIC, LoadsIC);
6476 }
6477
6478 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)
;
6479 return SmallIC;
6480 }
6481
6482 // Interleave if this is a large loop (small loops are already dealt with by
6483 // this point) that could benefit from interleaving.
6484 bool HasReductions = (Legal->getReductionVars()->size() > 0);
6485 if (TTI.enableAggressiveInterleaving(HasReductions)) {
6486 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)
;
6487 return IC;
6488 }
6489
6490 DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n"
; } } while (false)
;
6491 return 1;
6492}
6493
6494SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
6495LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
6496 // This function calculates the register usage by measuring the highest number
6497 // of values that are alive at a single location. Obviously, this is a very
6498 // rough estimation. We scan the loop in a topological order in order and
6499 // assign a number to each instruction. We use RPO to ensure that defs are
6500 // met before their users. We assume that each instruction that has in-loop
6501 // users starts an interval. We record every time that an in-loop value is
6502 // used, so we have a list of the first and last occurrences of each
6503 // instruction. Next, we transpose this data structure into a multi map that
6504 // holds the list of intervals that *end* at a specific location. This multi
6505 // map allows us to perform a linear search. We scan the instructions linearly
6506 // and record each time that a new interval starts, by placing it in a set.
6507 // If we find this value in the multi-map then we remove it from the set.
6508 // The max register usage is the maximum size of the set.
6509 // We also search for instructions that are defined outside the loop, but are
6510 // used inside the loop. We need this number separately from the max-interval
6511 // usage number because when we unroll, loop-invariant values do not take
6512 // more register.
6513 LoopBlocksDFS DFS(TheLoop);
6514 DFS.perform(LI);
6515
6516 RegisterUsage RU;
6517 RU.NumInstructions = 0;
6518
6519 // Each 'key' in the map opens a new interval. The values
6520 // of the map are the index of the 'last seen' usage of the
6521 // instruction that is the key.
6522 typedef DenseMap<Instruction *, unsigned> IntervalMap;
6523 // Maps instruction to its index.
6524 DenseMap<unsigned, Instruction *> IdxToInstr;
6525 // Marks the end of each interval.
6526 IntervalMap EndPoint;
6527 // Saves the list of instruction indices that are used in the loop.
6528 SmallSet<Instruction *, 8> Ends;
6529 // Saves the list of values that are used in the loop but are
6530 // defined outside the loop, such as arguments and constants.
6531 SmallPtrSet<Value *, 8> LoopInvariants;
6532
6533 unsigned Index = 0;
6534 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
6535 RU.NumInstructions += BB->size();
6536 for (Instruction &I : *BB) {
6537 IdxToInstr[Index++] = &I;
6538
6539 // Save the end location of each USE.
6540 for (Value *U : I.operands()) {
6541 auto *Instr = dyn_cast<Instruction>(U);
6542
6543 // Ignore non-instruction values such as arguments, constants, etc.
6544 if (!Instr)
6545 continue;
6546
6547 // If this instruction is outside the loop then record it and continue.
6548 if (!TheLoop->contains(Instr)) {
6549 LoopInvariants.insert(Instr);
6550 continue;
6551 }
6552
6553 // Overwrite previous end points.
6554 EndPoint[Instr] = Index;
6555 Ends.insert(Instr);
6556 }
6557 }
6558 }
6559
6560 // Saves the list of intervals that end with the index in 'key'.
6561 typedef SmallVector<Instruction *, 2> InstrList;
6562 DenseMap<unsigned, InstrList> TransposeEnds;
6563
6564 // Transpose the EndPoints to a list of values that end at each index.
6565 for (auto &Interval : EndPoint)
6566 TransposeEnds[Interval.second].push_back(Interval.first);
6567
6568 SmallSet<Instruction *, 8> OpenIntervals;
6569
6570 // Get the size of the widest register.
6571 unsigned MaxSafeDepDist = -1U;
6572 if (Legal->getMaxSafeDepDistBytes() != -1U)
6573 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
6574 unsigned WidestRegister =
6575 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
6576 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6577
6578 SmallVector<RegisterUsage, 8> RUs(VFs.size());
6579 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
6580
6581 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)
;
6582
6583 // A lambda that gets the register usage for the given type and VF.
6584 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
6585 if (Ty->isTokenTy())
6586 return 0U;
6587 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
6588 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
6589 };
6590
6591 for (unsigned int i = 0; i < Index; ++i) {
6592 Instruction *I = IdxToInstr[i];
6593
6594 // Remove all of the instructions that end at this location.
6595 InstrList &List = TransposeEnds[i];
6596 for (Instruction *ToRemove : List)
6597 OpenIntervals.erase(ToRemove);
6598
6599 // Ignore instructions that are never used within the loop.
6600 if (!Ends.count(I))
6601 continue;
6602
6603 // Skip ignored values.
6604 if (ValuesToIgnore.count(I))
6605 continue;
6606
6607 // For each VF find the maximum usage of registers.
6608 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
6609 if (VFs[j] == 1) {
6610 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
6611 continue;
6612 }
6613 collectUniformsAndScalars(VFs[j]);
6614 // Count the number of live intervals.
6615 unsigned RegUsage = 0;
6616 for (auto Inst : OpenIntervals) {
6617 // Skip ignored values for VF > 1.
6618 if (VecValuesToIgnore.count(Inst) ||
6619 isScalarAfterVectorization(Inst, VFs[j]))
6620 continue;
6621 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
6622 }
6623 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
6624 }
6625
6626 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)
6627 << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): At #" <<
i << " Interval # " << OpenIntervals.size() <<
'\n'; } } while (false)
;
6628
6629 // Add the current instruction to the list of open intervals.
6630 OpenIntervals.insert(I);
6631 }
6632
6633 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
6634 unsigned Invariant = 0;
6635 if (VFs[i] == 1)
6636 Invariant = LoopInvariants.size();
6637 else {
6638 for (auto Inst : LoopInvariants)
6639 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
6640 }
6641
6642 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)
;
6643 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)
;
6644 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)
;
6645 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)
;
6646
6647 RU.LoopInvariantRegs = Invariant;
6648 RU.MaxLocalUsers = MaxUsages[i];
6649 RUs[i] = RU;
6650 }
6651
6652 return RUs;
6653}
6654
6655void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
6656
6657 // If we aren't vectorizing the loop, or if we've already collected the
6658 // instructions to scalarize, there's nothing to do. Collection may already
6659 // have occurred if we have a user-selected VF and are now computing the
6660 // expected cost for interleaving.
6661 if (VF < 2 || InstsToScalarize.count(VF))
6662 return;
6663
6664 // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
6665 // not profitable to scalarize any instructions, the presence of VF in the
6666 // map will indicate that we've analyzed it already.
6667 ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
6668
6669 // Find all the instructions that are scalar with predication in the loop and
6670 // determine if it would be better to not if-convert the blocks they are in.
6671 // If so, we also record the instructions to scalarize.
6672 for (BasicBlock *BB : TheLoop->blocks()) {
6673 if (!Legal->blockNeedsPredication(BB))
6674 continue;
6675 for (Instruction &I : *BB)
6676 if (Legal->isScalarWithPredication(&I)) {
6677 ScalarCostsTy ScalarCosts;
6678 if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
6679 ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
6680 }
6681 }
6682}
6683
6684int LoopVectorizationCostModel::computePredInstDiscount(
6685 Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
6686 unsigned VF) {
6687
6688 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6689, __PRETTY_FUNCTION__))
6689 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6689, __PRETTY_FUNCTION__))
;
6690
6691 // Initialize the discount to zero, meaning that the scalar version and the
6692 // vector version cost the same.
6693 int Discount = 0;
6694
6695 // Holds instructions to analyze. The instructions we visit are mapped in
6696 // ScalarCosts. Those instructions are the ones that would be scalarized if
6697 // we find that the scalar version costs less.
6698 SmallVector<Instruction *, 8> Worklist;
6699
6700 // Returns true if the given instruction can be scalarized.
6701 auto canBeScalarized = [&](Instruction *I) -> bool {
6702
6703 // We only attempt to scalarize instructions forming a single-use chain
6704 // from the original predicated block that would otherwise be vectorized.
6705 // Although not strictly necessary, we give up on instructions we know will
6706 // already be scalar to avoid traversing chains that are unlikely to be
6707 // beneficial.
6708 if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
6709 isScalarAfterVectorization(I, VF))
6710 return false;
6711
6712 // If the instruction is scalar with predication, it will be analyzed
6713 // separately. We ignore it within the context of PredInst.
6714 if (Legal->isScalarWithPredication(I))
6715 return false;
6716
6717 // If any of the instruction's operands are uniform after vectorization,
6718 // the instruction cannot be scalarized. This prevents, for example, a
6719 // masked load from being scalarized.
6720 //
6721 // We assume we will only emit a value for lane zero of an instruction
6722 // marked uniform after vectorization, rather than VF identical values.
6723 // Thus, if we scalarize an instruction that uses a uniform, we would
6724 // create uses of values corresponding to the lanes we aren't emitting code
6725 // for. This behavior can be changed by allowing getScalarValue to clone
6726 // the lane zero values for uniforms rather than asserting.
6727 for (Use &U : I->operands())
6728 if (auto *J = dyn_cast<Instruction>(U.get()))
6729 if (isUniformAfterVectorization(J, VF))
6730 return false;
6731
6732 // Otherwise, we can scalarize the instruction.
6733 return true;
6734 };
6735
6736 // Returns true if an operand that cannot be scalarized must be extracted
6737 // from a vector. We will account for this scalarization overhead below. Note
6738 // that the non-void predicated instructions are placed in their own blocks,
6739 // and their return values are inserted into vectors. Thus, an extract would
6740 // still be required.
6741 auto needsExtract = [&](Instruction *I) -> bool {
6742 return TheLoop->contains(I) && !isScalarAfterVectorization(I, VF);
6743 };
6744
6745 // Compute the expected cost discount from scalarizing the entire expression
6746 // feeding the predicated instruction. We currently only consider expressions
6747 // that are single-use instruction chains.
6748 Worklist.push_back(PredInst);
6749 while (!Worklist.empty()) {
6750 Instruction *I = Worklist.pop_back_val();
6751
6752 // If we've already analyzed the instruction, there's nothing to do.
6753 if (ScalarCosts.count(I))
6754 continue;
6755
6756 // Compute the cost of the vector instruction. Note that this cost already
6757 // includes the scalarization overhead of the predicated instruction.
6758 unsigned VectorCost = getInstructionCost(I, VF).first;
6759
6760 // Compute the cost of the scalarized instruction. This cost is the cost of
6761 // the instruction as if it wasn't if-converted and instead remained in the
6762 // predicated block. We will scale this cost by block probability after
6763 // computing the scalarization overhead.
6764 unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
6765
6766 // Compute the scalarization overhead of needed insertelement instructions
6767 // and phi nodes.
6768 if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
6769 ScalarCost += TTI.getScalarizationOverhead(ToVectorTy(I->getType(), VF),
6770 true, false);
6771 ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
6772 }
6773
6774 // Compute the scalarization overhead of needed extractelement
6775 // instructions. For each of the instruction's operands, if the operand can
6776 // be scalarized, add it to the worklist; otherwise, account for the
6777 // overhead.
6778 for (Use &U : I->operands())
6779 if (auto *J = dyn_cast<Instruction>(U.get())) {
6780 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6781, __PRETTY_FUNCTION__))
6781 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6781, __PRETTY_FUNCTION__))
;
6782 if (canBeScalarized(J))
6783 Worklist.push_back(J);
6784 else if (needsExtract(J))
6785 ScalarCost += TTI.getScalarizationOverhead(
6786 ToVectorTy(J->getType(),VF), false, true);
6787 }
6788
6789 // Scale the total scalar cost by block probability.
6790 ScalarCost /= getReciprocalPredBlockProb();
6791
6792 // Compute the discount. A non-negative discount means the vector version
6793 // of the instruction costs more, and scalarizing would be beneficial.
6794 Discount += VectorCost - ScalarCost;
6795 ScalarCosts[I] = ScalarCost;
6796 }
6797
6798 return Discount;
6799}
6800
6801LoopVectorizationCostModel::VectorizationCostTy
6802LoopVectorizationCostModel::expectedCost(unsigned VF) {
6803 VectorizationCostTy Cost;
6804
6805 // Collect Uniform and Scalar instructions after vectorization with VF.
6806 collectUniformsAndScalars(VF);
6807
6808 // Collect the instructions (and their associated costs) that will be more
6809 // profitable to scalarize.
6810 collectInstsToScalarize(VF);
6811
6812 // For each block.
6813 for (BasicBlock *BB : TheLoop->blocks()) {
6814 VectorizationCostTy BlockCost;
6815
6816 // For each instruction in the old loop.
6817 for (Instruction &I : *BB) {
6818 // Skip dbg intrinsics.
6819 if (isa<DbgInfoIntrinsic>(I))
6820 continue;
6821
6822 // Skip ignored values.
6823 if (ValuesToIgnore.count(&I))
6824 continue;
6825
6826 VectorizationCostTy C = getInstructionCost(&I, VF);
6827
6828 // Check if we should override the cost.
6829 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
6830 C.first = ForceTargetInstructionCost;
6831
6832 BlockCost.first += C.first;
6833 BlockCost.second |= C.second;
6834 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)
6835 << 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)
;
6836 }
6837
6838 // If we are vectorizing a predicated block, it will have been
6839 // if-converted. This means that the block's instructions (aside from
6840 // stores and instructions that may divide by zero) will now be
6841 // unconditionally executed. For the scalar case, we may not always execute
6842 // the predicated block. Thus, scale the block's cost by the probability of
6843 // executing it.
6844 if (VF == 1 && Legal->blockNeedsPredication(BB))
6845 BlockCost.first /= getReciprocalPredBlockProb();
6846
6847 Cost.first += BlockCost.first;
6848 Cost.second |= BlockCost.second;
6849 }
6850
6851 return Cost;
6852}
6853
6854/// \brief Gets Address Access SCEV after verifying that the access pattern
6855/// is loop invariant except the induction variable dependence.
6856///
6857/// This SCEV can be sent to the Target in order to estimate the address
6858/// calculation cost.
6859static const SCEV *getAddressAccessSCEV(
6860 Value *Ptr,
6861 LoopVectorizationLegality *Legal,
6862 ScalarEvolution *SE,
6863 const Loop *TheLoop) {
6864 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
6865 if (!Gep)
6866 return nullptr;
6867
6868 // We are looking for a gep with all loop invariant indices except for one
6869 // which should be an induction variable.
6870 unsigned NumOperands = Gep->getNumOperands();
6871 for (unsigned i = 1; i < NumOperands; ++i) {
6872 Value *Opd = Gep->getOperand(i);
6873 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
6874 !Legal->isInductionVariable(Opd))
6875 return nullptr;
6876 }
6877
6878 // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.
6879 return SE->getSCEV(Ptr);
6880}
6881
6882static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
6883 return Legal->hasStride(I->getOperand(0)) ||
6884 Legal->hasStride(I->getOperand(1));
6885}
6886
6887unsigned LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,
6888 unsigned VF) {
6889 Type *ValTy = getMemInstValueType(I);
6890 auto SE = PSE.getSE();
6891
6892 unsigned Alignment = getMemInstAlignment(I);
6893 unsigned AS = getMemInstAddressSpace(I);
6894 Value *Ptr = getPointerOperand(I);
6895 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6896
6897 // Figure out whether the access is strided and get the stride value
6898 // if it's known in compile time
6899 const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop);
6900
6901 // Get the cost of the scalar memory instruction and address computation.
6902 unsigned Cost = VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
6903
6904 Cost += VF *
6905 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment,
6906 AS);
6907
6908 // Get the overhead of the extractelement and insertelement instructions
6909 // we might create due to scalarization.
6910 Cost += getScalarizationOverhead(I, VF, TTI);
6911
6912 // If we have a predicated store, it may not be executed for each vector
6913 // lane. Scale the cost by the probability of executing the predicated
6914 // block.
6915 if (Legal->isScalarWithPredication(I))
6916 Cost /= getReciprocalPredBlockProb();
6917
6918 return Cost;
6919}
6920
6921unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
6922 unsigned VF) {
6923 Type *ValTy = getMemInstValueType(I);
6924 Type *VectorTy = ToVectorTy(ValTy, VF);
6925 unsigned Alignment = getMemInstAlignment(I);
6926 Value *Ptr = getPointerOperand(I);
6927 unsigned AS = getMemInstAddressSpace(I);
6928 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6929
6930 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6931, __PRETTY_FUNCTION__))
6931 "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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6931, __PRETTY_FUNCTION__))
;
6932 unsigned Cost = 0;
6933 if (Legal->isMaskRequired(I))
6934 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6935 else
6936 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6937
6938 bool Reverse = ConsecutiveStride < 0;
6939 if (Reverse)
6940 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
6941 return Cost;
6942}
6943
6944unsigned LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I,
6945 unsigned VF) {
6946 LoadInst *LI = cast<LoadInst>(I);
6947 Type *ValTy = LI->getType();
6948 Type *VectorTy = ToVectorTy(ValTy, VF);
6949 unsigned Alignment = LI->getAlignment();
6950 unsigned AS = LI->getPointerAddressSpace();
6951
6952 return TTI.getAddressComputationCost(ValTy) +
6953 TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS) +
6954 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy);
6955}
6956
6957unsigned LoopVectorizationCostModel::getGatherScatterCost(Instruction *I,
6958 unsigned VF) {
6959 Type *ValTy = getMemInstValueType(I);
6960 Type *VectorTy = ToVectorTy(ValTy, VF);
6961 unsigned Alignment = getMemInstAlignment(I);
6962 Value *Ptr = getPointerOperand(I);
6963
6964 return TTI.getAddressComputationCost(VectorTy) +
6965 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
6966 Legal->isMaskRequired(I), Alignment);
6967}
6968
6969unsigned LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I,
6970 unsigned VF) {
6971 Type *ValTy = getMemInstValueType(I);
6972 Type *VectorTy = ToVectorTy(ValTy, VF);
6973 unsigned AS = getMemInstAddressSpace(I);
6974
6975 auto Group = Legal->getInterleavedAccessGroup(I);
6976 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6976, __PRETTY_FUNCTION__))
;
6977
6978 unsigned InterleaveFactor = Group->getFactor();
6979 Type *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor);
6980
6981 // Holds the indices of existing members in an interleaved load group.
6982 // An interleaved store group doesn't need this as it doesn't allow gaps.
6983 SmallVector<unsigned, 4> Indices;
6984 if (isa<LoadInst>(I)) {
6985 for (unsigned i = 0; i < InterleaveFactor; i++)
6986 if (Group->getMember(i))
6987 Indices.push_back(i);
6988 }
6989
6990 // Calculate the cost of the whole interleaved group.
6991 unsigned Cost = TTI.getInterleavedMemoryOpCost(I->getOpcode(), WideVecTy,
6992 Group->getFactor(), Indices,
6993 Group->getAlignment(), AS);
6994
6995 if (Group->isReverse())
6996 Cost += Group->getNumMembers() *
6997 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
6998 return Cost;
6999}
7000
7001unsigned LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I,
7002 unsigned VF) {
7003
7004 // Calculate scalar cost only. Vectorization cost should be ready at this
7005 // moment.
7006 if (VF == 1) {
7007 Type *ValTy = getMemInstValueType(I);
7008 unsigned Alignment = getMemInstAlignment(I);
7009 unsigned AS = getMemInstAlignment(I);
7010
7011 return TTI.getAddressComputationCost(ValTy) +
7012 TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS);
7013 }
7014 return getWideningCost(I, VF);
7015}
7016
7017LoopVectorizationCostModel::VectorizationCostTy
7018LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
7019 // If we know that this instruction will remain uniform, check the cost of
7020 // the scalar version.
7021 if (isUniformAfterVectorization(I, VF))
7022 VF = 1;
7023
7024 if (VF > 1 && isProfitableToScalarize(I, VF))
7025 return VectorizationCostTy(InstsToScalarize[VF][I], false);
7026
7027 Type *VectorTy;
7028 unsigned C = getInstructionCost(I, VF, VectorTy);
7029
7030 bool TypeNotScalarized =
7031 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
7032 return VectorizationCostTy(C, TypeNotScalarized);
7033}
7034
7035void LoopVectorizationCostModel::setCostBasedWideningDecision(unsigned VF) {
7036 if (VF == 1)
7037 return;
7038 for (BasicBlock *BB : TheLoop->blocks()) {
7039 // For each instruction in the old loop.
7040 for (Instruction &I : *BB) {
7041 Value *Ptr = getPointerOperand(&I);
7042 if (!Ptr)
7043 continue;
7044
7045 if (isa<LoadInst>(&I) && Legal->isUniform(Ptr)) {
7046 // Scalar load + broadcast
7047 unsigned Cost = getUniformMemOpCost(&I, VF);
7048 setWideningDecision(&I, VF, CM_Scalarize, Cost);
7049 continue;
7050 }
7051
7052 // We assume that widening is the best solution when possible.
7053 if (Legal->memoryInstructionCanBeWidened(&I, VF)) {
7054 unsigned Cost = getConsecutiveMemOpCost(&I, VF);
7055 setWideningDecision(&I, VF, CM_Widen, Cost);
7056 continue;
7057 }
7058
7059 // Choose between Interleaving, Gather/Scatter or Scalarization.
7060 unsigned InterleaveCost = UINT_MAX(2147483647 *2U +1U);
7061 unsigned NumAccesses = 1;
7062 if (Legal->isAccessInterleaved(&I)) {
7063 auto Group = Legal->getInterleavedAccessGroup(&I);
7064 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~svn295818/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7064, __PRETTY_FUNCTION__))
;
7065
7066 // Make one decision for the whole group.
7067 if (getWideningDecision(&I, VF) != CM_Unknown)
7068 continue;
7069
7070 NumAccesses = Group->getNumMembers();
7071 InterleaveCost = getInterleaveGroupCost(&I, VF);
7072 }
7073
7074 unsigned GatherScatterCost =
7075 Legal->isLegalGatherOrScatter(&I)
7076 ? getGatherScatterCost(&I, VF) * NumAccesses
7077 : UINT_MAX(2147483647 *2U +1U);
7078
7079 unsigned ScalarizationCost =
7080 getMemInstScalarizationCost(&I, VF) * NumAccesses;
7081
7082 // Choose better solution for the current VF,
7083 // write down this decision and use it during vectorization.
7084 unsigned Cost;
7085 InstWidening Decision;
7086 if (InterleaveCost <= GatherScatterCost &&
7087 InterleaveCost < ScalarizationCost) {
7088 Decision = CM_Interleave;
7089 Cost = InterleaveCost;
7090 } else if (GatherScatterCost < ScalarizationCost) {
7091 Decision = CM_GatherScatter;
7092 Cost = GatherScatterCost;
7093 } else {
7094 Decision = CM_Scalarize;
7095 Cost = ScalarizationCost;
7096 }
7097 // If the instructions belongs to an interleave group, the whole group
7098 // receives the same decision. The whole group receives the cost, but
7099 // the cost will actually be assigned to one instruction.
7100 if (auto Group = Legal->getInterleavedAccessGroup(&I))
7101 setWideningDecision(Group, VF, Decision, Cost);
7102 else
7103 setWideningDecision(&I, VF, Decision, Cost);
7104 }
7105 }
7106}
7107
7108unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
7109 unsigned VF,
7110 Type *&VectorTy) {
7111 Type *RetTy = I->getType();
7112 if (canTruncateToMinimalBitwidth(I, VF))
7113 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
7114 VectorTy = ToVectorTy(RetTy, VF);
7115 auto SE = PSE.getSE();
7116
7117 // TODO: We need to estimate the cost of intrinsic calls.
7118 switch (I->getOpcode()) {
7119 case Instruction::GetElementPtr:
7120 // We mark this instruction as zero-cost because the cost of GEPs in
7121 // vectorized code depends on whether the corresponding memory instruction
7122 // is scalarized or not. Therefore, we handle GEPs with the memory
7123 // instruction cost.
7124 return 0;
7125 case Instruction::Br: {
7126 return TTI.getCFInstrCost(I->getOpcode());
7127 }
7128 case Instruction::PHI: {
7129 auto *Phi = cast<PHINode>(I);
7130
7131 // First-order recurrences are replaced by vector shuffles inside the loop.
7132 if (VF > 1 && Legal->isFirstOrderRecurrence(Phi))
7133 return TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
7134 VectorTy, VF - 1, VectorTy);
7135
7136 // TODO: IF-converted IFs become selects.
7137 return 0;
7138 }
7139 case Instruction::UDiv:
7140 case Instruction::SDiv:
7141 case Instruction::URem:
7142 case Instruction::SRem:
7143 // If we have a predicated instruction, it may not be executed for each
7144 // vector lane. Get the scalarization cost and scale this amount by the
7145 // probability of executing the predicated block. If the instruction is not
7146 // predicated, we fall through to the next case.
7147 if (VF > 1 && Legal->isScalarWithPredication(I)) {
7148 unsigned Cost = 0;
7149
7150 // These instructions have a non-void type, so account for the phi nodes
7151 // that we will create. This cost is likely to be zero. The phi node
7152 // cost, if any, should be scaled by the block probability because it
7153 // models a copy at the end of each predicated block.
7154 Cost += VF * TTI.getCFInstrCost(Instruction::PHI);
7155
7156 // The cost of the non-predicated instruction.
7157 Cost += VF * TTI.getArithmeticInstrCost(I->getOpcode(), RetTy);
7158
7159 // The cost of insertelement and extractelement instructions needed for
7160 // scalarization.
7161 Cost += getScalarizationOverhead(I, VF, TTI);
7162
7163 // Scale the cost by the probability of executing the predicated blocks.
7164 // This assumes the predicated block for each vector lane is equally
7165 // likely.
7166 return Cost / getReciprocalPredBlockProb();
7167 }
7168 case Instruction::Add:
7169 case Instruction::FAdd:
7170 case Instruction::Sub:
7171 case Instruction::FSub:
7172 case Instruction::Mul:
7173 case Instruction::FMul:
7174 case Instruction::FDiv:
7175 case Instruction::FRem:
7176 case Instruction::Shl:
7177 case Instruction::LShr:
7178 case Instruction::AShr:
7179 case Instruction::And:
7180 case Instruction::Or:
7181 case Instruction::Xor: {
7182 // Since we will replace the stride by 1 the multiplication should go away.
7183 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
7184 return 0;
7185 // Certain instructions can be cheaper to vectorize if they have a constant
7186 // second vector operand. One example of this are shifts on x86.
7187 TargetTransformInfo::OperandValueKind Op1VK =
7188 TargetTransformInfo::OK_AnyValue;
7189 TargetTransformInfo::OperandValueKind Op2VK =
7190 TargetTransformInfo::OK_AnyValue;
7191 TargetTransformInfo::OperandValueProperties Op1VP =
7192 TargetTransformInfo::OP_None;
7193 TargetTransformInfo::OperandValueProperties Op2VP =
7194 TargetTransformInfo::OP_None;
7195 Value *Op2 = I->getOperand(1);
7196
7197 // Check for a splat or for a non uniform vector of constants.
7198 if (isa<ConstantInt>(Op2)) {
7199 ConstantInt *CInt = cast<ConstantInt>(Op2);
7200 if (CInt && CInt->getValue().isPowerOf2())
7201 Op2VP = TargetTransformInfo::OP_PowerOf2;
7202 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
7203 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
7204 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
7205 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
7206 if (SplatValue) {
7207 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
7208 if (CInt && CInt->getValue().isPowerOf2())
7209 Op2VP = TargetTransformInfo::OP_PowerOf2;
7210 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
7211 }
7212 } else if (Legal->isUniform(Op2)) {
7213 Op2VK = TargetTransformInfo::OK_UniformValue;
7214 }
7215 SmallVector<const Value *, 4> Operands(I->operand_values());
7216 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK,
7217 Op2VK, Op1VP, Op2VP, Operands);
7218 }
7219 case Instruction::Select: {
7220 SelectInst *SI = cast<SelectInst>(I);
7221 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
7222 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
7223 Type *CondTy = SI->getCondition()->getType();
7224 if (!ScalarCond)
7225 CondTy = VectorType::get(CondTy, VF);
7226
7227 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
7228 }
7229 case Instruction::ICmp:
7230 case Instruction::FCmp: {
7231 Type *ValTy = I->getOperand(0)->getType();
7232 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
7233 if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF))
7234 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]);
7235 VectorTy = ToVectorTy(ValTy, VF);
7236 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
7237 }
7238 case Instruction::Store:
7239 case Instruction::Load: {
7240 VectorTy = ToVectorTy(getMemInstValueType(I), VF);
7241 return getMemoryInstructionCost(I, VF);
7242 }
7243 case Instruction::ZExt:
7244 case Instruction::SExt:
7245 case Instruction::FPToUI:
7246 case Instruction::FPToSI:
7247 case Instruction::FPExt:
7248 case Instruction::PtrToInt:
7249 case Instruction::IntToPtr:
7250 case Instruction::SIToFP:
7251 case Instruction::UIToFP:
7252 case Instruction::Trunc:
7253 case Instruction::FPTrunc:
7254 case Instruction::BitCast: {
7255 // We optimize the truncation of induction variables having constant
7256 // integer steps. The cost of these truncations is the same as the scalar
7257 // operation.
7258 if (isOptimizableIVTruncate(I, VF)) {
7259 auto *Trunc = cast<TruncInst>(I);
7260 return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(),
7261 Trunc->getSrcTy());
7262 }
7263
7264 Type *SrcScalarTy = I->getOperand(0)->getType();
7265 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
7266 if (canTruncateToMinimalBitwidth(I, VF)) {
7267 // This cast is going to be shrunk. This may remove the cast or it might
7268 // turn it into slightly different cast. For example, if MinBW == 16,
7269 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
7270 //
7271 // Calculate the modified src and dest types.
7272 Type *MinVecTy = VectorTy;
7273 if (I->getOpcode() == Instruction::Trunc) {
7274 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
7275 VectorTy =
7276 largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7277 } else if (I->getOpcode() == Instruction::ZExt ||
7278 I->getOpcode() == Instruction::SExt) {
7279 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
7280 VectorTy =
7281 smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7282 }
7283 }
7284
7285 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
7286 }
7287 case Instruction::Call: {
7288 bool NeedToScalarize;
7289 CallInst *CI = cast<CallInst>(I);
7290 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
7291 if (getVectorIntrinsicIDForCall(CI, TLI))
7292 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
7293 return CallCost;
7294 }
7295 default:
7296 // The cost of executing VF copies of the scalar instruction. This opcode
7297 // is unknown. Assume that it is the same as 'mul'.
7298 return VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy) +
7299 getScalarizationOverhead(I, VF, TTI);
7300 } // end of switch.
7301}
7302
7303char LoopVectorize::ID = 0;
7304static const char lv_name[] = "Loop Vectorization";
7305INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)