Bug Summary

File:lib/Transforms/Vectorize/LoopVectorize.cpp
Warning:line 6621, column 60
Division by zero

Annotated Source Code

1//===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11// and generates target-independent LLVM-IR.
12// The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13// of instructions in order to estimate the profitability of vectorization.
14//
15// The loop vectorizer combines consecutive loop iterations into a single
16// 'wide' iteration. After this transformation the index is incremented
17// by the SIMD vector width, and not by one.
18//
19// This pass has three parts:
20// 1. The main loop pass that drives the different parts.
21// 2. LoopVectorizationLegality - A unit that checks for the legality
22// of the vectorization.
23// 3. InnerLoopVectorizer - A unit that performs the actual
24// widening of instructions.
25// 4. LoopVectorizationCostModel - A unit that checks for the profitability
26// of vectorization. It decides on the optimal vector width, which
27// can be one, if vectorization is not profitable.
28//
29//===----------------------------------------------------------------------===//
30//
31// The reduction-variable vectorization is based on the paper:
32// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33//
34// Variable uniformity checks are inspired by:
35// Karrenberg, R. and Hack, S. Whole Function Vectorization.
36//
37// The interleaved access vectorization is based on the paper:
38// Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
39// Data for SIMD
40//
41// Other ideas/concepts are from:
42// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
43//
44// S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45// Vectorizing Compilers.
46//
47//===----------------------------------------------------------------------===//
48
49#include "llvm/Transforms/Vectorize/LoopVectorize.h"
50#include "llvm/ADT/DenseMap.h"
51#include "llvm/ADT/Hashing.h"
52#include "llvm/ADT/MapVector.h"
53#include "llvm/ADT/Optional.h"
54#include "llvm/ADT/SCCIterator.h"
55#include "llvm/ADT/SetVector.h"
56#include "llvm/ADT/SmallPtrSet.h"
57#include "llvm/ADT/SmallSet.h"
58#include "llvm/ADT/SmallVector.h"
59#include "llvm/ADT/Statistic.h"
60#include "llvm/ADT/StringExtras.h"
61#include "llvm/Analysis/CodeMetrics.h"
62#include "llvm/Analysis/GlobalsModRef.h"
63#include "llvm/Analysis/LoopInfo.h"
64#include "llvm/Analysis/LoopIterator.h"
65#include "llvm/Analysis/LoopPass.h"
66#include "llvm/Analysis/ScalarEvolutionExpander.h"
67#include "llvm/Analysis/ScalarEvolutionExpressions.h"
68#include "llvm/Analysis/ValueTracking.h"
69#include "llvm/Analysis/VectorUtils.h"
70#include "llvm/IR/Constants.h"
71#include "llvm/IR/DataLayout.h"
72#include "llvm/IR/DebugInfo.h"
73#include "llvm/IR/DerivedTypes.h"
74#include "llvm/IR/DiagnosticInfo.h"
75#include "llvm/IR/Dominators.h"
76#include "llvm/IR/Function.h"
77#include "llvm/IR/IRBuilder.h"
78#include "llvm/IR/Instructions.h"
79#include "llvm/IR/IntrinsicInst.h"
80#include "llvm/IR/LLVMContext.h"
81#include "llvm/IR/Module.h"
82#include "llvm/IR/PatternMatch.h"
83#include "llvm/IR/Type.h"
84#include "llvm/IR/User.h"
85#include "llvm/IR/Value.h"
86#include "llvm/IR/ValueHandle.h"
87#include "llvm/IR/Verifier.h"
88#include "llvm/Pass.h"
89#include "llvm/Support/BranchProbability.h"
90#include "llvm/Support/CommandLine.h"
91#include "llvm/Support/Debug.h"
92#include "llvm/Support/raw_ostream.h"
93#include "llvm/Transforms/Scalar.h"
94#include "llvm/Transforms/Utils/BasicBlockUtils.h"
95#include "llvm/Transforms/Utils/Local.h"
96#include "llvm/Transforms/Utils/LoopSimplify.h"
97#include "llvm/Transforms/Utils/LoopUtils.h"
98#include "llvm/Transforms/Utils/LoopVersioning.h"
99#include "llvm/Transforms/Vectorize.h"
100#include <algorithm>
101#include <map>
102#include <tuple>
103
104using namespace llvm;
105using namespace llvm::PatternMatch;
106
107#define LV_NAME"loop-vectorize" "loop-vectorize"
108#define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize"
109
110STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized"
, "Number of loops vectorized", {0}, false}
;
111STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed"
, "Number of loops analyzed for vectorization", {0}, false}
;
112
113static cl::opt<bool>
114 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
115 cl::desc("Enable if-conversion during vectorization."));
116
117/// We don't vectorize loops with a known constant trip count below this number.
118static cl::opt<unsigned> TinyTripCountVectorThreshold(
119 "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
120 cl::desc("Don't vectorize loops with a constant "
121 "trip count that is smaller than this "
122 "value."));
123
124static cl::opt<bool> MaximizeBandwidth(
125 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
126 cl::desc("Maximize bandwidth when selecting vectorization factor which "
127 "will be determined by the smallest type in loop."));
128
129static cl::opt<bool> EnableInterleavedMemAccesses(
130 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
131 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
132
133/// Maximum factor for an interleaved memory access.
134static cl::opt<unsigned> MaxInterleaveGroupFactor(
135 "max-interleave-group-factor", cl::Hidden,
136 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
137 cl::init(8));
138
139/// We don't interleave loops with a known constant trip count below this
140/// number.
141static const unsigned TinyTripCountInterleaveThreshold = 128;
142
143static cl::opt<unsigned> ForceTargetNumScalarRegs(
144 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
145 cl::desc("A flag that overrides the target's number of scalar registers."));
146
147static cl::opt<unsigned> ForceTargetNumVectorRegs(
148 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
149 cl::desc("A flag that overrides the target's number of vector registers."));
150
151/// Maximum vectorization interleave count.
152static const unsigned MaxInterleaveFactor = 16;
153
154static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
155 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's max interleave factor for "
157 "scalar loops."));
158
159static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
160 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's max interleave factor for "
162 "vectorized loops."));
163
164static cl::opt<unsigned> ForceTargetInstructionCost(
165 "force-target-instruction-cost", cl::init(0), cl::Hidden,
166 cl::desc("A flag that overrides the target's expected cost for "
167 "an instruction to a single constant value. Mostly "
168 "useful for getting consistent testing."));
169
170static cl::opt<unsigned> SmallLoopCost(
171 "small-loop-cost", cl::init(20), cl::Hidden,
172 cl::desc(
173 "The cost of a loop that is considered 'small' by the interleaver."));
174
175static cl::opt<bool> LoopVectorizeWithBlockFrequency(
176 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
177 cl::desc("Enable the use of the block frequency analysis to access PGO "
178 "heuristics minimizing code growth in cold regions and being more "
179 "aggressive in hot regions."));
180
181// Runtime interleave loops for load/store throughput.
182static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
183 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
184 cl::desc(
185 "Enable runtime interleaving until load/store ports are saturated"));
186
187/// The number of stores in a loop that are allowed to need predication.
188static cl::opt<unsigned> NumberOfStoresToPredicate(
189 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
190 cl::desc("Max number of stores to be predicated behind an if."));
191
192static cl::opt<bool> EnableIndVarRegisterHeur(
193 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
194 cl::desc("Count the induction variable only once when interleaving"));
195
196static cl::opt<bool> EnableCondStoresVectorization(
197 "enable-cond-stores-vec", cl::init(true), cl::Hidden,
198 cl::desc("Enable if predication of stores during vectorization."));
199
200static cl::opt<unsigned> MaxNestedScalarReductionIC(
201 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
202 cl::desc("The maximum interleave count to use when interleaving a scalar "
203 "reduction in a nested loop."));
204
205static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
206 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
207 cl::desc("The maximum allowed number of runtime memory checks with a "
208 "vectorize(enable) pragma."));
209
210static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
211 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
212 cl::desc("The maximum number of SCEV checks allowed."));
213
214static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
215 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
216 cl::desc("The maximum number of SCEV checks allowed with a "
217 "vectorize(enable) pragma"));
218
219/// Create an analysis remark that explains why vectorization failed
220///
221/// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
222/// RemarkName is the identifier for the remark. If \p I is passed it is an
223/// instruction that prevents vectorization. Otherwise \p TheLoop is used for
224/// the location of the remark. \return the remark object that can be
225/// streamed to.
226static OptimizationRemarkAnalysis
227createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
228 Instruction *I = nullptr) {
229 Value *CodeRegion = TheLoop->getHeader();
230 DebugLoc DL = TheLoop->getStartLoc();
231
232 if (I) {
233 CodeRegion = I->getParent();
234 // If there is no debug location attached to the instruction, revert back to
235 // using the loop's.
236 if (I->getDebugLoc())
237 DL = I->getDebugLoc();
238 }
239
240 OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
241 R << "loop not vectorized: ";
242 return R;
243}
244
245namespace {
246
247// Forward declarations.
248class LoopVectorizeHints;
249class LoopVectorizationLegality;
250class LoopVectorizationCostModel;
251class LoopVectorizationRequirements;
252
253/// Returns true if the given loop body has a cycle, excluding the loop
254/// itself.
255static bool hasCyclesInLoopBody(const Loop &L) {
256 if (!L.empty())
257 return true;
258
259 for (const auto &SCC :
260 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
261 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
262 if (SCC.size() > 1) {
263 DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LVL: Detected a cycle in the loop body:\n"
; } } while (false)
;
264 DEBUG(L.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { L.dump(); } } while (false)
;
265 return true;
266 }
267 }
268 return false;
269}
270
271/// A helper function for converting Scalar types to vector types.
272/// If the incoming type is void, we return void. If the VF is 1, we return
273/// the scalar type.
274static Type *ToVectorTy(Type *Scalar, unsigned VF) {
275 if (Scalar->isVoidTy() || VF == 1)
276 return Scalar;
277 return VectorType::get(Scalar, VF);
278}
279
280// FIXME: The following helper functions have multiple implementations
281// in the project. They can be effectively organized in a common Load/Store
282// utilities unit.
283
284/// A helper function that returns the pointer operand of a load or store
285/// instruction.
286static Value *getPointerOperand(Value *I) {
287 if (auto *LI = dyn_cast<LoadInst>(I))
288 return LI->getPointerOperand();
289 if (auto *SI = dyn_cast<StoreInst>(I))
290 return SI->getPointerOperand();
291 return nullptr;
292}
293
294/// A helper function that returns the type of loaded or stored value.
295static Type *getMemInstValueType(Value *I) {
296 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 297, __PRETTY_FUNCTION__))
297 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 297, __PRETTY_FUNCTION__))
;
298 if (auto *LI = dyn_cast<LoadInst>(I))
299 return LI->getType();
300 return cast<StoreInst>(I)->getValueOperand()->getType();
301}
302
303/// A helper function that returns the alignment of load or store instruction.
304static unsigned getMemInstAlignment(Value *I) {
305 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 306, __PRETTY_FUNCTION__))
306 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 306, __PRETTY_FUNCTION__))
;
307 if (auto *LI = dyn_cast<LoadInst>(I))
308 return LI->getAlignment();
309 return cast<StoreInst>(I)->getAlignment();
310}
311
312/// A helper function that returns the address space of the pointer operand of
313/// load or store instruction.
314static unsigned getMemInstAddressSpace(Value *I) {
315 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 316, __PRETTY_FUNCTION__))
316 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 316, __PRETTY_FUNCTION__))
;
317 if (auto *LI = dyn_cast<LoadInst>(I))
318 return LI->getPointerAddressSpace();
319 return cast<StoreInst>(I)->getPointerAddressSpace();
320}
321
322/// A helper function that returns true if the given type is irregular. The
323/// type is irregular if its allocated size doesn't equal the store size of an
324/// element of the corresponding vector type at the given vectorization factor.
325static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
326
327 // Determine if an array of VF elements of type Ty is "bitcast compatible"
328 // with a <VF x Ty> vector.
329 if (VF > 1) {
330 auto *VectorTy = VectorType::get(Ty, VF);
331 return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
332 }
333
334 // If the vectorization factor is one, we just check if an array of type Ty
335 // requires padding between elements.
336 return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
337}
338
339/// A helper function that returns the reciprocal of the block probability of
340/// predicated blocks. If we return X, we are assuming the predicated block
341/// will execute once for for every X iterations of the loop header.
342///
343/// TODO: We should use actual block probability here, if available. Currently,
344/// we always assume predicated blocks have a 50% chance of executing.
345static unsigned getReciprocalPredBlockProb() { return 2; }
346
347/// A helper function that adds a 'fast' flag to floating-point operations.
348static Value *addFastMathFlag(Value *V) {
349 if (isa<FPMathOperator>(V)) {
350 FastMathFlags Flags;
351 Flags.setUnsafeAlgebra();
352 cast<Instruction>(V)->setFastMathFlags(Flags);
353 }
354 return V;
355}
356
357/// A helper function that returns an integer or floating-point constant with
358/// value C.
359static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) {
360 return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C)
361 : ConstantFP::get(Ty, C);
362}
363
364/// InnerLoopVectorizer vectorizes loops which contain only one basic
365/// block to a specified vectorization factor (VF).
366/// This class performs the widening of scalars into vectors, or multiple
367/// scalars. This class also implements the following features:
368/// * It inserts an epilogue loop for handling loops that don't have iteration
369/// counts that are known to be a multiple of the vectorization factor.
370/// * It handles the code generation for reduction variables.
371/// * Scalarization (implementation using scalars) of un-vectorizable
372/// instructions.
373/// InnerLoopVectorizer does not perform any vectorization-legality
374/// checks, and relies on the caller to check for the different legality
375/// aspects. The InnerLoopVectorizer relies on the
376/// LoopVectorizationLegality class to provide information about the induction
377/// and reduction variables that were found to a given vectorization factor.
378class InnerLoopVectorizer {
379public:
380 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
381 LoopInfo *LI, DominatorTree *DT,
382 const TargetLibraryInfo *TLI,
383 const TargetTransformInfo *TTI, AssumptionCache *AC,
384 OptimizationRemarkEmitter *ORE, unsigned VecWidth,
385 unsigned UnrollFactor, LoopVectorizationLegality *LVL,
386 LoopVectorizationCostModel *CM)
387 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
388 AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
389 Builder(PSE.getSE()->getContext()), Induction(nullptr),
390 OldInduction(nullptr), VectorLoopValueMap(UnrollFactor, VecWidth),
391 TripCount(nullptr), VectorTripCount(nullptr), Legal(LVL), Cost(CM),
392 AddedSafetyChecks(false) {}
393
394 // Perform the actual loop widening (vectorization).
395 void vectorize() {
396 // Create a new empty loop. Unlink the old loop and connect the new one.
397 createEmptyLoop();
398 // Widen each instruction in the old loop to a new one in the new loop.
399 vectorizeLoop();
400 }
401
402 // Return true if any runtime check is added.
403 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
404
405 virtual ~InnerLoopVectorizer() {}
406
407protected:
408 /// A small list of PHINodes.
409 typedef SmallVector<PHINode *, 4> PhiVector;
410
411 /// A type for vectorized values in the new loop. Each value from the
412 /// original loop, when vectorized, is represented by UF vector values in the
413 /// new unrolled loop, where UF is the unroll factor.
414 typedef SmallVector<Value *, 2> VectorParts;
415
416 /// A type for scalarized values in the new loop. Each value from the
417 /// original loop, when scalarized, is represented by UF x VF scalar values
418 /// in the new unrolled loop, where UF is the unroll factor and VF is the
419 /// vectorization factor.
420 typedef SmallVector<SmallVector<Value *, 4>, 2> ScalarParts;
421
422 // When we if-convert we need to create edge masks. We have to cache values
423 // so that we don't end up with exponential recursion/IR.
424 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
425 EdgeMaskCacheTy;
426 typedef DenseMap<BasicBlock *, VectorParts> BlockMaskCacheTy;
427
428 /// Create an empty loop, based on the loop ranges of the old loop.
429 void createEmptyLoop();
430
431 /// Set up the values of the IVs correctly when exiting the vector loop.
432 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
433 Value *CountRoundDown, Value *EndValue,
434 BasicBlock *MiddleBlock);
435
436 /// Create a new induction variable inside L.
437 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
438 Value *Step, Instruction *DL);
439 /// Copy and widen the instructions from the old loop.
440 virtual void vectorizeLoop();
441
442 /// Handle all cross-iteration phis in the header.
443 void fixCrossIterationPHIs();
444
445 /// Fix a first-order recurrence. This is the second phase of vectorizing
446 /// this phi node.
447 void fixFirstOrderRecurrence(PHINode *Phi);
448
449 /// Fix a reduction cross-iteration phi. This is the second phase of
450 /// vectorizing this phi node.
451 void fixReduction(PHINode *Phi);
452
453 /// \brief The Loop exit block may have single value PHI nodes where the
454 /// incoming value is 'Undef'. While vectorizing we only handled real values
455 /// that were defined inside the loop. Here we fix the 'undef case'.
456 /// See PR14725.
457 void fixLCSSAPHIs();
458
459 /// Iteratively sink the scalarized operands of a predicated instruction into
460 /// the block that was created for it.
461 void sinkScalarOperands(Instruction *PredInst);
462
463 /// Predicate conditional instructions that require predication on their
464 /// respective conditions.
465 void predicateInstructions();
466
467 /// Collect the instructions from the original loop that would be trivially
468 /// dead in the vectorized loop if generated.
469 void collectTriviallyDeadInstructions(
470 SmallPtrSetImpl<Instruction *> &DeadInstructions);
471
472 /// Shrinks vector element sizes to the smallest bitwidth they can be legally
473 /// represented as.
474 void truncateToMinimalBitwidths();
475
476 /// A helper function that computes the predicate of the block BB, assuming
477 /// that the header block of the loop is set to True. It returns the *entry*
478 /// mask for the block BB.
479 VectorParts createBlockInMask(BasicBlock *BB);
480 /// A helper function that computes the predicate of the edge between SRC
481 /// and DST.
482 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
483
484 /// A helper function to vectorize a single instruction within the innermost
485 /// loop.
486 void vectorizeInstruction(Instruction &I);
487
488 /// Vectorize a single PHINode in a block. This method handles the induction
489 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
490 /// arbitrary length vectors.
491 void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF);
492
493 /// Insert the new loop to the loop hierarchy and pass manager
494 /// and update the analysis passes.
495 void updateAnalysis();
496
497 /// This instruction is un-vectorizable. Implement it as a sequence
498 /// of scalars. If \p IfPredicateInstr is true we need to 'hide' each
499 /// scalarized instruction behind an if block predicated on the control
500 /// dependence of the instruction.
501 virtual void scalarizeInstruction(Instruction *Instr,
502 bool IfPredicateInstr = false);
503
504 /// Vectorize Load and Store instructions,
505 virtual void vectorizeMemoryInstruction(Instruction *Instr);
506
507 /// Create a broadcast instruction. This method generates a broadcast
508 /// instruction (shuffle) for loop invariant values and for the induction
509 /// value. If this is the induction variable then we extend it to N, N+1, ...
510 /// this is needed because each iteration in the loop corresponds to a SIMD
511 /// element.
512 virtual Value *getBroadcastInstrs(Value *V);
513
514 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
515 /// to each vector element of Val. The sequence starts at StartIndex.
516 /// \p Opcode is relevant for FP induction variable.
517 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
518 Instruction::BinaryOps Opcode =
519 Instruction::BinaryOpsEnd);
520
521 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
522 /// variable on which to base the steps, \p Step is the size of the step, and
523 /// \p EntryVal is the value from the original loop that maps to the steps.
524 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
525 /// can be a truncate instruction).
526 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal,
527 const InductionDescriptor &ID);
528
529 /// Create a vector induction phi node based on an existing scalar one. \p
530 /// EntryVal is the value from the original loop that maps to the vector phi
531 /// node, and \p Step is the loop-invariant step. If \p EntryVal is a
532 /// truncate instruction, instead of widening the original IV, we widen a
533 /// version of the IV truncated to \p EntryVal's type.
534 void createVectorIntOrFpInductionPHI(const InductionDescriptor &II,
535 Value *Step, Instruction *EntryVal);
536
537 /// Widen an integer or floating-point induction variable \p IV. If \p Trunc
538 /// is provided, the integer induction variable will first be truncated to
539 /// the corresponding type.
540 void widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc = nullptr);
541
542 /// Returns true if an instruction \p I should be scalarized instead of
543 /// vectorized for the chosen vectorization factor.
544 bool shouldScalarizeInstruction(Instruction *I) const;
545
546 /// Returns true if we should generate a scalar version of \p IV.
547 bool needsScalarInduction(Instruction *IV) const;
548
549 /// Return a constant reference to the VectorParts corresponding to \p V from
550 /// the original loop. If the value has already been vectorized, the
551 /// corresponding vector entry in VectorLoopValueMap is returned. If,
552 /// however, the value has a scalar entry in VectorLoopValueMap, we construct
553 /// new vector values on-demand by inserting the scalar values into vectors
554 /// with an insertelement sequence. If the value has been neither vectorized
555 /// nor scalarized, it must be loop invariant, so we simply broadcast the
556 /// value into vectors.
557 const VectorParts &getVectorValue(Value *V);
558
559 /// Return a value in the new loop corresponding to \p V from the original
560 /// loop at unroll index \p Part and vector index \p Lane. If the value has
561 /// been vectorized but not scalarized, the necessary extractelement
562 /// instruction will be generated.
563 Value *getScalarValue(Value *V, unsigned Part, unsigned Lane);
564
565 /// Try to vectorize the interleaved access group that \p Instr belongs to.
566 void vectorizeInterleaveGroup(Instruction *Instr);
567
568 /// Generate a shuffle sequence that will reverse the vector Vec.
569 virtual Value *reverseVector(Value *Vec);
570
571 /// Returns (and creates if needed) the original loop trip count.
572 Value *getOrCreateTripCount(Loop *NewLoop);
573
574 /// Returns (and creates if needed) the trip count of the widened loop.
575 Value *getOrCreateVectorTripCount(Loop *NewLoop);
576
577 /// Emit a bypass check to see if the trip count would overflow, or we
578 /// wouldn't have enough iterations to execute one vector loop.
579 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
580 /// Emit a bypass check to see if the vector trip count is nonzero.
581 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
582 /// Emit a bypass check to see if all of the SCEV assumptions we've
583 /// had to make are correct.
584 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
585 /// Emit bypass checks to check any memory assumptions we may have made.
586 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
587
588 /// Add additional metadata to \p To that was not present on \p Orig.
589 ///
590 /// Currently this is used to add the noalias annotations based on the
591 /// inserted memchecks. Use this for instructions that are *cloned* into the
592 /// vector loop.
593 void addNewMetadata(Instruction *To, const Instruction *Orig);
594
595 /// Add metadata from one instruction to another.
596 ///
597 /// This includes both the original MDs from \p From and additional ones (\see
598 /// addNewMetadata). Use this for *newly created* instructions in the vector
599 /// loop.
600 void addMetadata(Instruction *To, Instruction *From);
601
602 /// \brief Similar to the previous function but it adds the metadata to a
603 /// vector of instructions.
604 void addMetadata(ArrayRef<Value *> To, Instruction *From);
605
606 /// \brief Set the debug location in the builder using the debug location in
607 /// the instruction.
608 void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr);
609
610 /// This is a helper class for maintaining vectorization state. It's used for
611 /// mapping values from the original loop to their corresponding values in
612 /// the new loop. Two mappings are maintained: one for vectorized values and
613 /// one for scalarized values. Vectorized values are represented with UF
614 /// vector values in the new loop, and scalarized values are represented with
615 /// UF x VF scalar values in the new loop. UF and VF are the unroll and
616 /// vectorization factors, respectively.
617 ///
618 /// Entries can be added to either map with initVector and initScalar, which
619 /// initialize and return a constant reference to the new entry. If a
620 /// non-constant reference to a vector entry is required, getVector can be
621 /// used to retrieve a mutable entry. We currently directly modify the mapped
622 /// values during "fix-up" operations that occur once the first phase of
623 /// widening is complete. These operations include type truncation and the
624 /// second phase of recurrence widening.
625 ///
626 /// Otherwise, entries from either map should be accessed using the
627 /// getVectorValue or getScalarValue functions from InnerLoopVectorizer.
628 /// getVectorValue and getScalarValue coordinate to generate a vector or
629 /// scalar value on-demand if one is not yet available. When vectorizing a
630 /// loop, we visit the definition of an instruction before its uses. When
631 /// visiting the definition, we either vectorize or scalarize the
632 /// instruction, creating an entry for it in the corresponding map. (In some
633 /// cases, such as induction variables, we will create both vector and scalar
634 /// entries.) Then, as we encounter uses of the definition, we derive values
635 /// for each scalar or vector use unless such a value is already available.
636 /// For example, if we scalarize a definition and one of its uses is vector,
637 /// we build the required vector on-demand with an insertelement sequence
638 /// when visiting the use. Otherwise, if the use is scalar, we can use the
639 /// existing scalar definition.
640 struct ValueMap {
641
642 /// Construct an empty map with the given unroll and vectorization factors.
643 ValueMap(unsigned UnrollFactor, unsigned VecWidth)
644 : UF(UnrollFactor), VF(VecWidth) {
645 // The unroll and vectorization factors are only used in asserts builds
646 // to verify map entries are sized appropriately.
647 (void)UF;
648 (void)VF;
649 }
650
651 /// \return True if the map has a vector entry for \p Key.
652 bool hasVector(Value *Key) const { return VectorMapStorage.count(Key); }
653
654 /// \return True if the map has a scalar entry for \p Key.
655 bool hasScalar(Value *Key) const { return ScalarMapStorage.count(Key); }
656
657 /// \brief Map \p Key to the given VectorParts \p Entry, and return a
658 /// constant reference to the new vector map entry. The given key should
659 /// not already be in the map, and the given VectorParts should be
660 /// correctly sized for the current unroll factor.
661 const VectorParts &initVector(Value *Key, const VectorParts &Entry) {
662 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 662, __PRETTY_FUNCTION__))
;
663 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 663, __PRETTY_FUNCTION__))
;
664 VectorMapStorage[Key] = Entry;
665 return VectorMapStorage[Key];
666 }
667
668 /// \brief Map \p Key to the given ScalarParts \p Entry, and return a
669 /// constant reference to the new scalar map entry. The given key should
670 /// not already be in the map, and the given ScalarParts should be
671 /// correctly sized for the current unroll and vectorization factors.
672 const ScalarParts &initScalar(Value *Key, const ScalarParts &Entry) {
673 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 673, __PRETTY_FUNCTION__))
;
674 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 679, __PRETTY_FUNCTION__))
675 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 679, __PRETTY_FUNCTION__))
676 [&](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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 679, __PRETTY_FUNCTION__))
677 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 679, __PRETTY_FUNCTION__))
678 }) &&((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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 679, __PRETTY_FUNCTION__))
679 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 679, __PRETTY_FUNCTION__))
;
680 ScalarMapStorage[Key] = Entry;
681 return ScalarMapStorage[Key];
682 }
683
684 /// \return A reference to the vector map entry corresponding to \p Key.
685 /// The key should already be in the map. This function should only be used
686 /// when it's necessary to update values that have already been vectorized.
687 /// This is the case for "fix-up" operations including type truncation and
688 /// the second phase of recurrence vectorization. If a non-const reference
689 /// isn't required, getVectorValue should be used instead.
690 VectorParts &getVector(Value *Key) {
691 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 691, __PRETTY_FUNCTION__))
;
692 return VectorMapStorage.find(Key)->second;
693 }
694
695 /// Retrieve an entry from the vector or scalar maps. The preferred way to
696 /// access an existing mapped entry is with getVectorValue or
697 /// getScalarValue from InnerLoopVectorizer. Until those functions can be
698 /// moved inside ValueMap, we have to declare them as friends.
699 friend const VectorParts &InnerLoopVectorizer::getVectorValue(Value *V);
700 friend Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
701 unsigned Lane);
702
703 private:
704 /// The unroll factor. Each entry in the vector map contains UF vector
705 /// values.
706 unsigned UF;
707
708 /// The vectorization factor. Each entry in the scalar map contains UF x VF
709 /// scalar values.
710 unsigned VF;
711
712 /// The vector and scalar map storage. We use std::map and not DenseMap
713 /// because insertions to DenseMap invalidate its iterators.
714 std::map<Value *, VectorParts> VectorMapStorage;
715 std::map<Value *, ScalarParts> ScalarMapStorage;
716 };
717
718 /// The original loop.
719 Loop *OrigLoop;
720 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
721 /// dynamic knowledge to simplify SCEV expressions and converts them to a
722 /// more usable form.
723 PredicatedScalarEvolution &PSE;
724 /// Loop Info.
725 LoopInfo *LI;
726 /// Dominator Tree.
727 DominatorTree *DT;
728 /// Alias Analysis.
729 AliasAnalysis *AA;
730 /// Target Library Info.
731 const TargetLibraryInfo *TLI;
732 /// Target Transform Info.
733 const TargetTransformInfo *TTI;
734 /// Assumption Cache.
735 AssumptionCache *AC;
736 /// Interface to emit optimization remarks.
737 OptimizationRemarkEmitter *ORE;
738
739 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
740 /// used.
741 ///
742 /// This is currently only used to add no-alias metadata based on the
743 /// memchecks. The actually versioning is performed manually.
744 std::unique_ptr<LoopVersioning> LVer;
745
746 /// The vectorization SIMD factor to use. Each vector will have this many
747 /// vector elements.
748 unsigned VF;
749
750protected:
751 /// The vectorization unroll factor to use. Each scalar is vectorized to this
752 /// many different vector instructions.
753 unsigned UF;
754
755 /// The builder that we use
756 IRBuilder<> Builder;
757
758 // --- Vectorization state ---
759
760 /// The vector-loop preheader.
761 BasicBlock *LoopVectorPreHeader;
762 /// The scalar-loop preheader.
763 BasicBlock *LoopScalarPreHeader;
764 /// Middle Block between the vector and the scalar.
765 BasicBlock *LoopMiddleBlock;
766 /// The ExitBlock of the scalar loop.
767 BasicBlock *LoopExitBlock;
768 /// The vector loop body.
769 BasicBlock *LoopVectorBody;
770 /// The scalar loop body.
771 BasicBlock *LoopScalarBody;
772 /// A list of all bypass blocks. The first block is the entry of the loop.
773 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
774
775 /// The new Induction variable which was added to the new block.
776 PHINode *Induction;
777 /// The induction variable of the old basic block.
778 PHINode *OldInduction;
779
780 /// Maps values from the original loop to their corresponding values in the
781 /// vectorized loop. A key value can map to either vector values, scalar
782 /// values or both kinds of values, depending on whether the key was
783 /// vectorized and scalarized.
784 ValueMap VectorLoopValueMap;
785
786 /// Store instructions that should be predicated, as a pair
787 /// <StoreInst, Predicate>
788 SmallVector<std::pair<Instruction *, Value *>, 4> PredicatedInstructions;
789 EdgeMaskCacheTy EdgeMaskCache;
790 BlockMaskCacheTy BlockMaskCache;
791 /// Trip count of the original loop.
792 Value *TripCount;
793 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
794 Value *VectorTripCount;
795
796 /// The legality analysis.
797 LoopVectorizationLegality *Legal;
798
799 /// The profitablity analysis.
800 LoopVectorizationCostModel *Cost;
801
802 // Record whether runtime checks are added.
803 bool AddedSafetyChecks;
804
805 // Holds the end values for each induction variable. We save the end values
806 // so we can later fix-up the external users of the induction variables.
807 DenseMap<PHINode *, Value *> IVEndValues;
808};
809
810class InnerLoopUnroller : public InnerLoopVectorizer {
811public:
812 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
813 LoopInfo *LI, DominatorTree *DT,
814 const TargetLibraryInfo *TLI,
815 const TargetTransformInfo *TTI, AssumptionCache *AC,
816 OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
817 LoopVectorizationLegality *LVL,
818 LoopVectorizationCostModel *CM)
819 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
820 UnrollFactor, LVL, CM) {}
821
822private:
823 void vectorizeMemoryInstruction(Instruction *Instr) override;
824 Value *getBroadcastInstrs(Value *V) override;
825 Value *getStepVector(Value *Val, int StartIdx, Value *Step,
826 Instruction::BinaryOps Opcode =
827 Instruction::BinaryOpsEnd) override;
828 Value *reverseVector(Value *Vec) override;
829};
830
831/// \brief Look for a meaningful debug location on the instruction or it's
832/// operands.
833static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
834 if (!I)
835 return I;
836
837 DebugLoc Empty;
838 if (I->getDebugLoc() != Empty)
839 return I;
840
841 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
842 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
843 if (OpInst->getDebugLoc() != Empty)
844 return OpInst;
845 }
846
847 return I;
848}
849
850void InnerLoopVectorizer::setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
851 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) {
852 const DILocation *DIL = Inst->getDebugLoc();
853 if (DIL && Inst->getFunction()->isDebugInfoForProfiling())
854 B.SetCurrentDebugLocation(DIL->cloneWithDuplicationFactor(UF * VF));
855 else
856 B.SetCurrentDebugLocation(DIL);
857 } else
858 B.SetCurrentDebugLocation(DebugLoc());
859}
860
861#ifndef NDEBUG
862/// \return string containing a file name and a line # for the given loop.
863static std::string getDebugLocString(const Loop *L) {
864 std::string Result;
865 if (L) {
866 raw_string_ostream OS(Result);
867 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
868 LoopDbgLoc.print(OS);
869 else
870 // Just print the module name.
871 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
872 OS.flush();
873 }
874 return Result;
875}
876#endif
877
878void InnerLoopVectorizer::addNewMetadata(Instruction *To,
879 const Instruction *Orig) {
880 // If the loop was versioned with memchecks, add the corresponding no-alias
881 // metadata.
882 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
883 LVer->annotateInstWithNoAlias(To, Orig);
884}
885
886void InnerLoopVectorizer::addMetadata(Instruction *To,
887 Instruction *From) {
888 propagateMetadata(To, From);
889 addNewMetadata(To, From);
890}
891
892void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
893 Instruction *From) {
894 for (Value *V : To) {
895 if (Instruction *I = dyn_cast<Instruction>(V))
896 addMetadata(I, From);
897 }
898}
899
900/// \brief The group of interleaved loads/stores sharing the same stride and
901/// close to each other.
902///
903/// Each member in this group has an index starting from 0, and the largest
904/// index should be less than interleaved factor, which is equal to the absolute
905/// value of the access's stride.
906///
907/// E.g. An interleaved load group of factor 4:
908/// for (unsigned i = 0; i < 1024; i+=4) {
909/// a = A[i]; // Member of index 0
910/// b = A[i+1]; // Member of index 1
911/// d = A[i+3]; // Member of index 3
912/// ...
913/// }
914///
915/// An interleaved store group of factor 4:
916/// for (unsigned i = 0; i < 1024; i+=4) {
917/// ...
918/// A[i] = a; // Member of index 0
919/// A[i+1] = b; // Member of index 1
920/// A[i+2] = c; // Member of index 2
921/// A[i+3] = d; // Member of index 3
922/// }
923///
924/// Note: the interleaved load group could have gaps (missing members), but
925/// the interleaved store group doesn't allow gaps.
926class InterleaveGroup {
927public:
928 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
929 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
930 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 930, __PRETTY_FUNCTION__))
;
931
932 Factor = std::abs(Stride);
933 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 933, __PRETTY_FUNCTION__))
;
934
935 Reverse = Stride < 0;
936 Members[0] = Instr;
937 }
938
939 bool isReverse() const { return Reverse; }
940 unsigned getFactor() const { return Factor; }
941 unsigned getAlignment() const { return Align; }
942 unsigned getNumMembers() const { return Members.size(); }
943
944 /// \brief Try to insert a new member \p Instr with index \p Index and
945 /// alignment \p NewAlign. The index is related to the leader and it could be
946 /// negative if it is the new leader.
947 ///
948 /// \returns false if the instruction doesn't belong to the group.
949 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
950 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 950, __PRETTY_FUNCTION__))
;
951
952 int Key = Index + SmallestKey;
953
954 // Skip if there is already a member with the same index.
955 if (Members.count(Key))
956 return false;
957
958 if (Key > LargestKey) {
959 // The largest index is always less than the interleave factor.
960 if (Index >= static_cast<int>(Factor))
961 return false;
962
963 LargestKey = Key;
964 } else if (Key < SmallestKey) {
965 // The largest index is always less than the interleave factor.
966 if (LargestKey - Key >= static_cast<int>(Factor))
967 return false;
968
969 SmallestKey = Key;
970 }
971
972 // It's always safe to select the minimum alignment.
973 Align = std::min(Align, NewAlign);
974 Members[Key] = Instr;
975 return true;
976 }
977
978 /// \brief Get the member with the given index \p Index
979 ///
980 /// \returns nullptr if contains no such member.
981 Instruction *getMember(unsigned Index) const {
982 int Key = SmallestKey + Index;
983 if (!Members.count(Key))
984 return nullptr;
985
986 return Members.find(Key)->second;
987 }
988
989 /// \brief Get the index for the given member. Unlike the key in the member
990 /// map, the index starts from 0.
991 unsigned getIndex(Instruction *Instr) const {
992 for (auto I : Members)
993 if (I.second == Instr)
994 return I.first - SmallestKey;
995
996 llvm_unreachable("InterleaveGroup contains no such member")::llvm::llvm_unreachable_internal("InterleaveGroup contains no such member"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 996)
;
997 }
998
999 Instruction *getInsertPos() const { return InsertPos; }
1000 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
1001
1002private:
1003 unsigned Factor; // Interleave Factor.
1004 bool Reverse;
1005 unsigned Align;
1006 DenseMap<int, Instruction *> Members;
1007 int SmallestKey;
1008 int LargestKey;
1009
1010 // To avoid breaking dependences, vectorized instructions of an interleave
1011 // group should be inserted at either the first load or the last store in
1012 // program order.
1013 //
1014 // E.g. %even = load i32 // Insert Position
1015 // %add = add i32 %even // Use of %even
1016 // %odd = load i32
1017 //
1018 // store i32 %even
1019 // %odd = add i32 // Def of %odd
1020 // store i32 %odd // Insert Position
1021 Instruction *InsertPos;
1022};
1023
1024/// \brief Drive the analysis of interleaved memory accesses in the loop.
1025///
1026/// Use this class to analyze interleaved accesses only when we can vectorize
1027/// a loop. Otherwise it's meaningless to do analysis as the vectorization
1028/// on interleaved accesses is unsafe.
1029///
1030/// The analysis collects interleave groups and records the relationships
1031/// between the member and the group in a map.
1032class InterleavedAccessInfo {
1033public:
1034 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1035 DominatorTree *DT, LoopInfo *LI)
1036 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
1037 RequiresScalarEpilogue(false) {}
1038
1039 ~InterleavedAccessInfo() {
1040 SmallSet<InterleaveGroup *, 4> DelSet;
1041 // Avoid releasing a pointer twice.
1042 for (auto &I : InterleaveGroupMap)
1043 DelSet.insert(I.second);
1044 for (auto *Ptr : DelSet)
1045 delete Ptr;
1046 }
1047
1048 /// \brief Analyze the interleaved accesses and collect them in interleave
1049 /// groups. Substitute symbolic strides using \p Strides.
1050 void analyzeInterleaving(const ValueToValueMap &Strides);
1051
1052 /// \brief Check if \p Instr belongs to any interleave group.
1053 bool isInterleaved(Instruction *Instr) const {
1054 return InterleaveGroupMap.count(Instr);
1055 }
1056
1057 /// \brief Return the maximum interleave factor of all interleaved groups.
1058 unsigned getMaxInterleaveFactor() const {
1059 unsigned MaxFactor = 1;
1060 for (auto &Entry : InterleaveGroupMap)
1061 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
1062 return MaxFactor;
1063 }
1064
1065 /// \brief Get the interleave group that \p Instr belongs to.
1066 ///
1067 /// \returns nullptr if doesn't have such group.
1068 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1069 if (InterleaveGroupMap.count(Instr))
1070 return InterleaveGroupMap.find(Instr)->second;
1071 return nullptr;
1072 }
1073
1074 /// \brief Returns true if an interleaved group that may access memory
1075 /// out-of-bounds requires a scalar epilogue iteration for correctness.
1076 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1077
1078 /// \brief Initialize the LoopAccessInfo used for dependence checking.
1079 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1080
1081private:
1082 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1083 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1084 /// The interleaved access analysis can also add new predicates (for example
1085 /// by versioning strides of pointers).
1086 PredicatedScalarEvolution &PSE;
1087 Loop *TheLoop;
1088 DominatorTree *DT;
1089 LoopInfo *LI;
1090 const LoopAccessInfo *LAI;
1091
1092 /// True if the loop may contain non-reversed interleaved groups with
1093 /// out-of-bounds accesses. We ensure we don't speculatively access memory
1094 /// out-of-bounds by executing at least one scalar epilogue iteration.
1095 bool RequiresScalarEpilogue;
1096
1097 /// Holds the relationships between the members and the interleave group.
1098 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1099
1100 /// Holds dependences among the memory accesses in the loop. It maps a source
1101 /// access to a set of dependent sink accesses.
1102 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
1103
1104 /// \brief The descriptor for a strided memory access.
1105 struct StrideDescriptor {
1106 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1107 unsigned Align)
1108 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1109
1110 StrideDescriptor() = default;
1111
1112 // The access's stride. It is negative for a reverse access.
1113 int64_t Stride = 0;
1114 const SCEV *Scev = nullptr; // The scalar expression of this access
1115 uint64_t Size = 0; // The size of the memory object.
1116 unsigned Align = 0; // The alignment of this access.
1117 };
1118
1119 /// \brief A type for holding instructions and their stride descriptors.
1120 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
1121
1122 /// \brief Create a new interleave group with the given instruction \p Instr,
1123 /// stride \p Stride and alignment \p Align.
1124 ///
1125 /// \returns the newly created interleave group.
1126 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1127 unsigned Align) {
1128 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1129, __PRETTY_FUNCTION__))
1129 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1129, __PRETTY_FUNCTION__))
;
1130 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1131 return InterleaveGroupMap[Instr];
1132 }
1133
1134 /// \brief Release the group and remove all the relationships.
1135 void releaseGroup(InterleaveGroup *Group) {
1136 for (unsigned i = 0; i < Group->getFactor(); i++)
1137 if (Instruction *Member = Group->getMember(i))
1138 InterleaveGroupMap.erase(Member);
1139
1140 delete Group;
1141 }
1142
1143 /// \brief Collect all the accesses with a constant stride in program order.
1144 void collectConstStrideAccesses(
1145 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1146 const ValueToValueMap &Strides);
1147
1148 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1149 static bool isStrided(int Stride) {
1150 unsigned Factor = std::abs(Stride);
1151 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1152 }
1153
1154 /// \brief Returns true if \p BB is a predicated block.
1155 bool isPredicated(BasicBlock *BB) const {
1156 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1157 }
1158
1159 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1160 bool areDependencesValid() const {
1161 return LAI && LAI->getDepChecker().getDependences();
1162 }
1163
1164 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1165 /// necessary, when constructing interleaved groups.
1166 ///
1167 /// \p A must precede \p B in program order. We return false if reordering is
1168 /// not necessary or is prevented because \p A and \p B may be dependent.
1169 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1170 StrideEntry *B) const {
1171
1172 // Code motion for interleaved accesses can potentially hoist strided loads
1173 // and sink strided stores. The code below checks the legality of the
1174 // following two conditions:
1175 //
1176 // 1. Potentially moving a strided load (B) before any store (A) that
1177 // precedes B, or
1178 //
1179 // 2. Potentially moving a strided store (A) after any load or store (B)
1180 // that A precedes.
1181 //
1182 // It's legal to reorder A and B if we know there isn't a dependence from A
1183 // to B. Note that this determination is conservative since some
1184 // dependences could potentially be reordered safely.
1185
1186 // A is potentially the source of a dependence.
1187 auto *Src = A->first;
1188 auto SrcDes = A->second;
1189
1190 // B is potentially the sink of a dependence.
1191 auto *Sink = B->first;
1192 auto SinkDes = B->second;
1193
1194 // Code motion for interleaved accesses can't violate WAR dependences.
1195 // Thus, reordering is legal if the source isn't a write.
1196 if (!Src->mayWriteToMemory())
1197 return true;
1198
1199 // At least one of the accesses must be strided.
1200 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1201 return true;
1202
1203 // If dependence information is not available from LoopAccessInfo,
1204 // conservatively assume the instructions can't be reordered.
1205 if (!areDependencesValid())
1206 return false;
1207
1208 // If we know there is a dependence from source to sink, assume the
1209 // instructions can't be reordered. Otherwise, reordering is legal.
1210 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1211 }
1212
1213 /// \brief Collect the dependences from LoopAccessInfo.
1214 ///
1215 /// We process the dependences once during the interleaved access analysis to
1216 /// enable constant-time dependence queries.
1217 void collectDependences() {
1218 if (!areDependencesValid())
1219 return;
1220 auto *Deps = LAI->getDepChecker().getDependences();
1221 for (auto Dep : *Deps)
1222 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1223 }
1224};
1225
1226/// Utility class for getting and setting loop vectorizer hints in the form
1227/// of loop metadata.
1228/// This class keeps a number of loop annotations locally (as member variables)
1229/// and can, upon request, write them back as metadata on the loop. It will
1230/// initially scan the loop for existing metadata, and will update the local
1231/// values based on information in the loop.
1232/// We cannot write all values to metadata, as the mere presence of some info,
1233/// for example 'force', means a decision has been made. So, we need to be
1234/// careful NOT to add them if the user hasn't specifically asked so.
1235class LoopVectorizeHints {
1236 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1237
1238 /// Hint - associates name and validation with the hint value.
1239 struct Hint {
1240 const char *Name;
1241 unsigned Value; // This may have to change for non-numeric values.
1242 HintKind Kind;
1243
1244 Hint(const char *Name, unsigned Value, HintKind Kind)
1245 : Name(Name), Value(Value), Kind(Kind) {}
1246
1247 bool validate(unsigned Val) {
1248 switch (Kind) {
1249 case HK_WIDTH:
1250 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1251 case HK_UNROLL:
1252 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1253 case HK_FORCE:
1254 return (Val <= 1);
1255 }
1256 return false;
1257 }
1258 };
1259
1260 /// Vectorization width.
1261 Hint Width;
1262 /// Vectorization interleave factor.
1263 Hint Interleave;
1264 /// Vectorization forced
1265 Hint Force;
1266
1267 /// Return the loop metadata prefix.
1268 static StringRef Prefix() { return "llvm.loop."; }
1269
1270 /// True if there is any unsafe math in the loop.
1271 bool PotentiallyUnsafe;
1272
1273public:
1274 enum ForceKind {
1275 FK_Undefined = -1, ///< Not selected.
1276 FK_Disabled = 0, ///< Forcing disabled.
1277 FK_Enabled = 1, ///< Forcing enabled.
1278 };
1279
1280 LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1281 OptimizationRemarkEmitter &ORE)
1282 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1283 HK_WIDTH),
1284 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1285 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1286 PotentiallyUnsafe(false), TheLoop(L), ORE(ORE) {
1287 // Populate values with existing loop metadata.
1288 getHintsFromMetadata();
1289
1290 // force-vector-interleave overrides DisableInterleaving.
1291 if (VectorizerParams::isInterleaveForced())
1292 Interleave.Value = VectorizerParams::VectorizationInterleave;
1293
1294 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)
1295 << "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)
;
1296 }
1297
1298 /// Mark the loop L as already vectorized by setting the width to 1.
1299 void setAlreadyVectorized() {
1300 Width.Value = Interleave.Value = 1;
1301 Hint Hints[] = {Width, Interleave};
1302 writeHintsToMetadata(Hints);
1303 }
1304
1305 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1306 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1307 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)
;
1308 emitRemarkWithHints();
1309 return false;
1310 }
1311
1312 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1313 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)
;
1314 emitRemarkWithHints();
1315 return false;
1316 }
1317
1318 if (getWidth() == 1 && getInterleave() == 1) {
1319 // FIXME: Add a separate metadata to indicate when the loop has already
1320 // been vectorized instead of setting width and count to 1.
1321 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)
;
1322 // FIXME: Add interleave.disable metadata. This will allow
1323 // vectorize.disable to be used without disabling the pass and errors
1324 // to differentiate between disabled vectorization and a width of 1.
1325 ORE.emit(OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1326 "AllDisabled", L->getStartLoc(),
1327 L->getHeader())
1328 << "loop not vectorized: vectorization and interleaving are "
1329 "explicitly disabled, or vectorize width and interleave "
1330 "count are both set to 1");
1331 return false;
1332 }
1333
1334 return true;
1335 }
1336
1337 /// Dumps all the hint information.
1338 void emitRemarkWithHints() const {
1339 using namespace ore;
1340 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1341 ORE.emit(OptimizationRemarkMissed(LV_NAME"loop-vectorize", "MissedExplicitlyDisabled",
1342 TheLoop->getStartLoc(),
1343 TheLoop->getHeader())
1344 << "loop not vectorized: vectorization is explicitly disabled");
1345 else {
1346 OptimizationRemarkMissed R(LV_NAME"loop-vectorize", "MissedDetails",
1347 TheLoop->getStartLoc(), TheLoop->getHeader());
1348 R << "loop not vectorized";
1349 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1350 R << " (Force=" << NV("Force", true);
1351 if (Width.Value != 0)
1352 R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1353 if (Interleave.Value != 0)
1354 R << ", Interleave Count=" << NV("InterleaveCount", Interleave.Value);
1355 R << ")";
1356 }
1357 ORE.emit(R);
1358 }
1359 }
1360
1361 unsigned getWidth() const { return Width.Value; }
1362 unsigned getInterleave() const { return Interleave.Value; }
1363 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1364
1365 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1366 /// pass name to force the frontend to print the diagnostic.
1367 const char *vectorizeAnalysisPassName() const {
1368 if (getWidth() == 1)
1369 return LV_NAME"loop-vectorize";
1370 if (getForce() == LoopVectorizeHints::FK_Disabled)
1371 return LV_NAME"loop-vectorize";
1372 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1373 return LV_NAME"loop-vectorize";
1374 return OptimizationRemarkAnalysis::AlwaysPrint;
1375 }
1376
1377 bool allowReordering() const {
1378 // When enabling loop hints are provided we allow the vectorizer to change
1379 // the order of operations that is given by the scalar loop. This is not
1380 // enabled by default because can be unsafe or inefficient. For example,
1381 // reordering floating-point operations will change the way round-off
1382 // error accumulates in the loop.
1383 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1384 }
1385
1386 bool isPotentiallyUnsafe() const {
1387 // Avoid FP vectorization if the target is unsure about proper support.
1388 // This may be related to the SIMD unit in the target not handling
1389 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1390 // Otherwise, a sequence of vectorized loops, even without reduction,
1391 // could lead to different end results on the destination vectors.
1392 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1393 }
1394
1395 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1396
1397private:
1398 /// Find hints specified in the loop metadata and update local values.
1399 void getHintsFromMetadata() {
1400 MDNode *LoopID = TheLoop->getLoopID();
1401 if (!LoopID)
1402 return;
1403
1404 // First operand should refer to the loop id itself.
1405 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1405, __PRETTY_FUNCTION__))
;
1406 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1406, __PRETTY_FUNCTION__))
;
1407
1408 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1409 const MDString *S = nullptr;
1410 SmallVector<Metadata *, 4> Args;
1411
1412 // The expected hint is either a MDString or a MDNode with the first
1413 // operand a MDString.
1414 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1415 if (!MD || MD->getNumOperands() == 0)
1416 continue;
1417 S = dyn_cast<MDString>(MD->getOperand(0));
1418 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1419 Args.push_back(MD->getOperand(i));
1420 } else {
1421 S = dyn_cast<MDString>(LoopID->getOperand(i));
1422 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1422, __PRETTY_FUNCTION__))
;
1423 }
1424
1425 if (!S)
1426 continue;
1427
1428 // Check if the hint starts with the loop metadata prefix.
1429 StringRef Name = S->getString();
1430 if (Args.size() == 1)
1431 setHint(Name, Args[0]);
1432 }
1433 }
1434
1435 /// Checks string hint with one operand and set value if valid.
1436 void setHint(StringRef Name, Metadata *Arg) {
1437 if (!Name.startswith(Prefix()))
1438 return;
1439 Name = Name.substr(Prefix().size(), StringRef::npos);
1440
1441 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1442 if (!C)
1443 return;
1444 unsigned Val = C->getZExtValue();
1445
1446 Hint *Hints[] = {&Width, &Interleave, &Force};
1447 for (auto H : Hints) {
1448 if (Name == H->Name) {
1449 if (H->validate(Val))
1450 H->Value = Val;
1451 else
1452 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)
;
1453 break;
1454 }
1455 }
1456 }
1457
1458 /// Create a new hint from name / value pair.
1459 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1460 LLVMContext &Context = TheLoop->getHeader()->getContext();
1461 Metadata *MDs[] = {MDString::get(Context, Name),
1462 ConstantAsMetadata::get(
1463 ConstantInt::get(Type::getInt32Ty(Context), V))};
1464 return MDNode::get(Context, MDs);
1465 }
1466
1467 /// Matches metadata with hint name.
1468 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1469 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1470 if (!Name)
1471 return false;
1472
1473 for (auto H : HintTypes)
1474 if (Name->getString().endswith(H.Name))
1475 return true;
1476 return false;
1477 }
1478
1479 /// Sets current hints into loop metadata, keeping other values intact.
1480 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1481 if (HintTypes.size() == 0)
1482 return;
1483
1484 // Reserve the first element to LoopID (see below).
1485 SmallVector<Metadata *, 4> MDs(1);
1486 // If the loop already has metadata, then ignore the existing operands.
1487 MDNode *LoopID = TheLoop->getLoopID();
1488 if (LoopID) {
1489 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1490 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1491 // If node in update list, ignore old value.
1492 if (!matchesHintMetadataName(Node, HintTypes))
1493 MDs.push_back(Node);
1494 }
1495 }
1496
1497 // Now, add the missing hints.
1498 for (auto H : HintTypes)
1499 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1500
1501 // Replace current metadata node with new one.
1502 LLVMContext &Context = TheLoop->getHeader()->getContext();
1503 MDNode *NewLoopID = MDNode::get(Context, MDs);
1504 // Set operand 0 to refer to the loop id itself.
1505 NewLoopID->replaceOperandWith(0, NewLoopID);
1506
1507 TheLoop->setLoopID(NewLoopID);
1508 }
1509
1510 /// The loop these hints belong to.
1511 const Loop *TheLoop;
1512
1513 /// Interface to emit optimization remarks.
1514 OptimizationRemarkEmitter &ORE;
1515};
1516
1517static void emitMissedWarning(Function *F, Loop *L,
1518 const LoopVectorizeHints &LH,
1519 OptimizationRemarkEmitter *ORE) {
1520 LH.emitRemarkWithHints();
1521
1522 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1523 if (LH.getWidth() != 1)
1524 ORE->emit(DiagnosticInfoOptimizationFailure(
1525 DEBUG_TYPE"loop-vectorize", "FailedRequestedVectorization",
1526 L->getStartLoc(), L->getHeader())
1527 << "loop not vectorized: "
1528 << "failed explicitly specified loop vectorization");
1529 else if (LH.getInterleave() != 1)
1530 ORE->emit(DiagnosticInfoOptimizationFailure(
1531 DEBUG_TYPE"loop-vectorize", "FailedRequestedInterleaving", L->getStartLoc(),
1532 L->getHeader())
1533 << "loop not interleaved: "
1534 << "failed explicitly specified loop interleaving");
1535 }
1536}
1537
1538/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1539/// to what vectorization factor.
1540/// This class does not look at the profitability of vectorization, only the
1541/// legality. This class has two main kinds of checks:
1542/// * Memory checks - The code in canVectorizeMemory checks if vectorization
1543/// will change the order of memory accesses in a way that will change the
1544/// correctness of the program.
1545/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1546/// checks for a number of different conditions, such as the availability of a
1547/// single induction variable, that all types are supported and vectorize-able,
1548/// etc. This code reflects the capabilities of InnerLoopVectorizer.
1549/// This class is also used by InnerLoopVectorizer for identifying
1550/// induction variable and the different reduction variables.
1551class LoopVectorizationLegality {
1552public:
1553 LoopVectorizationLegality(
1554 Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
1555 TargetLibraryInfo *TLI, AliasAnalysis *AA, Function *F,
1556 const TargetTransformInfo *TTI,
1557 std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1558 OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R,
1559 LoopVectorizeHints *H)
1560 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TTI(TTI), DT(DT),
1561 GetLAA(GetLAA), LAI(nullptr), ORE(ORE), InterleaveInfo(PSE, L, DT, LI),
1562 PrimaryInduction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1563 Requirements(R), Hints(H) {}
1564
1565 /// ReductionList contains the reduction descriptors for all
1566 /// of the reductions that were found in the loop.
1567 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1568
1569 /// InductionList saves induction variables and maps them to the
1570 /// induction descriptor.
1571 typedef MapVector<PHINode *, InductionDescriptor> InductionList;
1572
1573 /// RecurrenceSet contains the phi nodes that are recurrences other than
1574 /// inductions and reductions.
1575 typedef SmallPtrSet<const PHINode *, 8> RecurrenceSet;
1576
1577 /// Returns true if it is legal to vectorize this loop.
1578 /// This does not mean that it is profitable to vectorize this
1579 /// loop, only that it is legal to do so.
1580 bool canVectorize();
1581
1582 /// Returns the primary induction variable.
1583 PHINode *getPrimaryInduction() { return PrimaryInduction; }
1584
1585 /// Returns the reduction variables found in the loop.
1586 ReductionList *getReductionVars() { return &Reductions; }
1587
1588 /// Returns the induction variables found in the loop.
1589 InductionList *getInductionVars() { return &Inductions; }
1590
1591 /// Return the first-order recurrences found in the loop.
1592 RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1593
1594 /// Returns the widest induction type.
1595 Type *getWidestInductionType() { return WidestIndTy; }
1596
1597 /// Returns True if V is an induction variable in this loop.
1598 bool isInductionVariable(const Value *V);
1599
1600 /// Returns True if PN is a reduction variable in this loop.
1601 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1602
1603 /// Returns True if Phi is a first-order recurrence in this loop.
1604 bool isFirstOrderRecurrence(const PHINode *Phi);
1605
1606 /// Return true if the block BB needs to be predicated in order for the loop
1607 /// to be vectorized.
1608 bool blockNeedsPredication(BasicBlock *BB);
1609
1610 /// Check if this pointer is consecutive when vectorizing. This happens
1611 /// when the last index of the GEP is the induction variable, or that the
1612 /// pointer itself is an induction variable.
1613 /// This check allows us to vectorize A[idx] into a wide load/store.
1614 /// Returns:
1615 /// 0 - Stride is unknown or non-consecutive.
1616 /// 1 - Address is consecutive.
1617 /// -1 - Address is consecutive, and decreasing.
1618 int isConsecutivePtr(Value *Ptr);
1619
1620 /// Returns true if the value V is uniform within the loop.
1621 bool isUniform(Value *V);
1622
1623 /// Returns the information that we collected about runtime memory check.
1624 const RuntimePointerChecking *getRuntimePointerChecking() const {
1625 return LAI->getRuntimePointerChecking();
1626 }
1627
1628 const LoopAccessInfo *getLAI() const { return LAI; }
1629
1630 /// \brief Check if \p Instr belongs to any interleaved access group.
1631 bool isAccessInterleaved(Instruction *Instr) {
1632 return InterleaveInfo.isInterleaved(Instr);
1633 }
1634
1635 /// \brief Return the maximum interleave factor of all interleaved groups.
1636 unsigned getMaxInterleaveFactor() const {
1637 return InterleaveInfo.getMaxInterleaveFactor();
1638 }
1639
1640 /// \brief Get the interleaved access group that \p Instr belongs to.
1641 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1642 return InterleaveInfo.getInterleaveGroup(Instr);
1643 }
1644
1645 /// \brief Returns true if an interleaved group requires a scalar iteration
1646 /// to handle accesses with gaps.
1647 bool requiresScalarEpilogue() const {
1648 return InterleaveInfo.requiresScalarEpilogue();
1649 }
1650
1651 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1652
1653 bool hasStride(Value *V) { return LAI->hasStride(V); }
1654
1655 /// Returns true if the target machine supports masked store operation
1656 /// for the given \p DataType and kind of access to \p Ptr.
1657 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1658 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1659 }
1660 /// Returns true if the target machine supports masked load operation
1661 /// for the given \p DataType and kind of access to \p Ptr.
1662 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1663 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1664 }
1665 /// Returns true if the target machine supports masked scatter operation
1666 /// for the given \p DataType.
1667 bool isLegalMaskedScatter(Type *DataType) {
1668 return TTI->isLegalMaskedScatter(DataType);
1669 }
1670 /// Returns true if the target machine supports masked gather operation
1671 /// for the given \p DataType.
1672 bool isLegalMaskedGather(Type *DataType) {
1673 return TTI->isLegalMaskedGather(DataType);
1674 }
1675 /// Returns true if the target machine can represent \p V as a masked gather
1676 /// or scatter operation.
1677 bool isLegalGatherOrScatter(Value *V) {
1678 auto *LI = dyn_cast<LoadInst>(V);
1679 auto *SI = dyn_cast<StoreInst>(V);
1680 if (!LI && !SI)
1681 return false;
1682 auto *Ptr = getPointerOperand(V);
1683 auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1684 return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1685 }
1686
1687 /// Returns true if vector representation of the instruction \p I
1688 /// requires mask.
1689 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1690 unsigned getNumStores() const { return LAI->getNumStores(); }
1691 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1692 unsigned getNumPredStores() const { return NumPredStores; }
1693
1694 /// Returns true if \p I is an instruction that will be scalarized with
1695 /// predication. Such instructions include conditional stores and
1696 /// instructions that may divide by zero.
1697 bool isScalarWithPredication(Instruction *I);
1698
1699 /// Returns true if \p I is a memory instruction with consecutive memory
1700 /// access that can be widened.
1701 bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
1702
1703private:
1704 /// Check if a single basic block loop is vectorizable.
1705 /// At this point we know that this is a loop with a constant trip count
1706 /// and we only need to check individual instructions.
1707 bool canVectorizeInstrs();
1708
1709 /// When we vectorize loops we may change the order in which
1710 /// we read and write from memory. This method checks if it is
1711 /// legal to vectorize the code, considering only memory constrains.
1712 /// Returns true if the loop is vectorizable
1713 bool canVectorizeMemory();
1714
1715 /// Return true if we can vectorize this loop using the IF-conversion
1716 /// transformation.
1717 bool canVectorizeWithIfConvert();
1718
1719 /// Return true if all of the instructions in the block can be speculatively
1720 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1721 /// and we know that we can read from them without segfault.
1722 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1723
1724 /// Updates the vectorization state by adding \p Phi to the inductions list.
1725 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1726 /// better choice for the main induction than the existing one.
1727 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1728 SmallPtrSetImpl<Value *> &AllowedExit);
1729
1730 /// Create an analysis remark that explains why vectorization failed
1731 ///
1732 /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1733 /// an instruction that prevents vectorization. Otherwise the loop is used
1734 /// for the location of the remark. \return the remark object that can be
1735 /// streamed to.
1736 OptimizationRemarkAnalysis
1737 createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1738 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1739 RemarkName, TheLoop, I);
1740 }
1741
1742 /// \brief If an access has a symbolic strides, this maps the pointer value to
1743 /// the stride symbol.
1744 const ValueToValueMap *getSymbolicStrides() {
1745 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1746 // it's collected. This happens from canVectorizeWithIfConvert, when the
1747 // pointer is checked to reference consecutive elements suitable for a
1748 // masked access.
1749 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1750 }
1751
1752 unsigned NumPredStores;
1753
1754 /// The loop that we evaluate.
1755 Loop *TheLoop;
1756 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1757 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1758 /// of existing SCEV assumptions. The analysis will also add a minimal set
1759 /// of new predicates if this is required to enable vectorization and
1760 /// unrolling.
1761 PredicatedScalarEvolution &PSE;
1762 /// Target Library Info.
1763 TargetLibraryInfo *TLI;
1764 /// Target Transform Info
1765 const TargetTransformInfo *TTI;
1766 /// Dominator Tree.
1767 DominatorTree *DT;
1768 // LoopAccess analysis.
1769 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1770 // And the loop-accesses info corresponding to this loop. This pointer is
1771 // null until canVectorizeMemory sets it up.
1772 const LoopAccessInfo *LAI;
1773 /// Interface to emit optimization remarks.
1774 OptimizationRemarkEmitter *ORE;
1775
1776 /// The interleave access information contains groups of interleaved accesses
1777 /// with the same stride and close to each other.
1778 InterleavedAccessInfo InterleaveInfo;
1779
1780 // --- vectorization state --- //
1781
1782 /// Holds the primary induction variable. This is the counter of the
1783 /// loop.
1784 PHINode *PrimaryInduction;
1785 /// Holds the reduction variables.
1786 ReductionList Reductions;
1787 /// Holds all of the induction variables that we found in the loop.
1788 /// Notice that inductions don't need to start at zero and that induction
1789 /// variables can be pointers.
1790 InductionList Inductions;
1791 /// Holds the phi nodes that are first-order recurrences.
1792 RecurrenceSet FirstOrderRecurrences;
1793 /// Holds the widest induction type encountered.
1794 Type *WidestIndTy;
1795
1796 /// Allowed outside users. This holds the induction and reduction
1797 /// vars which can be accessed from outside the loop.
1798 SmallPtrSet<Value *, 4> AllowedExit;
1799
1800 /// Can we assume the absence of NaNs.
1801 bool HasFunNoNaNAttr;
1802
1803 /// Vectorization requirements that will go through late-evaluation.
1804 LoopVectorizationRequirements *Requirements;
1805
1806 /// Used to emit an analysis of any legality issues.
1807 LoopVectorizeHints *Hints;
1808
1809 /// While vectorizing these instructions we have to generate a
1810 /// call to the appropriate masked intrinsic
1811 SmallPtrSet<const Instruction *, 8> MaskedOp;
1812};
1813
1814/// LoopVectorizationCostModel - estimates the expected speedups due to
1815/// vectorization.
1816/// In many cases vectorization is not profitable. This can happen because of
1817/// a number of reasons. In this class we mainly attempt to predict the
1818/// expected speedup/slowdowns due to the supported instruction set. We use the
1819/// TargetTransformInfo to query the different backends for the cost of
1820/// different operations.
1821class LoopVectorizationCostModel {
1822public:
1823 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1824 LoopInfo *LI, LoopVectorizationLegality *Legal,
1825 const TargetTransformInfo &TTI,
1826 const TargetLibraryInfo *TLI, DemandedBits *DB,
1827 AssumptionCache *AC,
1828 OptimizationRemarkEmitter *ORE, const Function *F,
1829 const LoopVectorizeHints *Hints)
1830 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1831 AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1832
1833 /// \return An upper bound for the vectorization factor, or None if
1834 /// vectorization should be avoided up front.
1835 Optional<unsigned> computeMaxVF(bool OptForSize);
1836
1837 /// Information about vectorization costs
1838 struct VectorizationFactor {
1839 unsigned Width; // Vector width with best cost
1840 unsigned Cost; // Cost of the loop with that width
1841 };
1842 /// \return The most profitable vectorization factor and the cost of that VF.
1843 /// This method checks every power of two up to MaxVF. If UserVF is not ZERO
1844 /// then this vectorization factor will be selected if vectorization is
1845 /// possible.
1846 VectorizationFactor selectVectorizationFactor(unsigned MaxVF);
1847
1848 /// Setup cost-based decisions for user vectorization factor.
1849 void selectUserVectorizationFactor(unsigned UserVF) {
1850 collectUniformsAndScalars(UserVF);
1851 collectInstsToScalarize(UserVF);
1852 }
1853
1854 /// \return The size (in bits) of the smallest and widest types in the code
1855 /// that needs to be vectorized. We ignore values that remain scalar such as
1856 /// 64 bit loop indices.
1857 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1858
1859 /// \return The desired interleave count.
1860 /// If interleave count has been specified by metadata it will be returned.
1861 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1862 /// are the selected vectorization factor and the cost of the selected VF.
1863 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1864 unsigned LoopCost);
1865
1866 /// Memory access instruction may be vectorized in more than one way.
1867 /// Form of instruction after vectorization depends on cost.
1868 /// This function takes cost-based decisions for Load/Store instructions
1869 /// and collects them in a map. This decisions map is used for building
1870 /// the lists of loop-uniform and loop-scalar instructions.
1871 /// The calculated cost is saved with widening decision in order to
1872 /// avoid redundant calculations.
1873 void setCostBasedWideningDecision(unsigned VF);
1874
1875 /// \brief A struct that represents some properties of the register usage
1876 /// of a loop.
1877 struct RegisterUsage {
1878 /// Holds the number of loop invariant values that are used in the loop.
1879 unsigned LoopInvariantRegs;
1880 /// Holds the maximum number of concurrent live intervals in the loop.
1881 unsigned MaxLocalUsers;
1882 /// Holds the number of instructions in the loop.
1883 unsigned NumInstructions;
1884 };
1885
1886 /// \return Returns information about the register usages of the loop for the
1887 /// given vectorization factors.
1888 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1889
1890 /// Collect values we want to ignore in the cost model.
1891 void collectValuesToIgnore();
1892
1893 /// \returns The smallest bitwidth each instruction can be represented with.
1894 /// The vector equivalents of these instructions should be truncated to this
1895 /// type.
1896 const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1897 return MinBWs;
1898 }
1899
1900 /// \returns True if it is more profitable to scalarize instruction \p I for
1901 /// vectorization factor \p VF.
1902 bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1903 auto Scalars = InstsToScalarize.find(VF);
1904 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1905, __PRETTY_FUNCTION__))
1905 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1905, __PRETTY_FUNCTION__))
;
1906 return Scalars->second.count(I);
1907 }
1908
1909 /// Returns true if \p I is known to be uniform after vectorization.
1910 bool isUniformAfterVectorization(Instruction *I, unsigned VF) const {
1911 if (VF == 1)
1912 return true;
1913 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1913, __PRETTY_FUNCTION__))
;
1914 auto UniformsPerVF = Uniforms.find(VF);
1915 return UniformsPerVF->second.count(I);
1916 }
1917
1918 /// Returns true if \p I is known to be scalar after vectorization.
1919 bool isScalarAfterVectorization(Instruction *I, unsigned VF) const {
1920 if (VF == 1)
1921 return true;
1922 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1922, __PRETTY_FUNCTION__))
;
1923 auto ScalarsPerVF = Scalars.find(VF);
1924 return ScalarsPerVF->second.count(I);
1925 }
1926
1927 /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1928 /// for vectorization factor \p VF.
1929 bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1930 return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1931 !isScalarAfterVectorization(I, VF);
1932 }
1933
1934 /// Decision that was taken during cost calculation for memory instruction.
1935 enum InstWidening {
1936 CM_Unknown,
1937 CM_Widen,
1938 CM_Interleave,
1939 CM_GatherScatter,
1940 CM_Scalarize
1941 };
1942
1943 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1944 /// instruction \p I and vector width \p VF.
1945 void setWideningDecision(Instruction *I, unsigned VF, InstWidening W,
1946 unsigned Cost) {
1947 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1947, __PRETTY_FUNCTION__))
;
1948 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1949 }
1950
1951 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1952 /// interleaving group \p Grp and vector width \p VF.
1953 void setWideningDecision(const InterleaveGroup *Grp, unsigned VF,
1954 InstWidening W, unsigned Cost) {
1955 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1955, __PRETTY_FUNCTION__))
;
1956 /// Broadcast this decicion to all instructions inside the group.
1957 /// But the cost will be assigned to one instruction only.
1958 for (unsigned i = 0; i < Grp->getFactor(); ++i) {
1959 if (auto *I = Grp->getMember(i)) {
1960 if (Grp->getInsertPos() == I)
1961 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1962 else
1963 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0);
1964 }
1965 }
1966 }
1967
1968 /// Return the cost model decision for the given instruction \p I and vector
1969 /// width \p VF. Return CM_Unknown if this instruction did not pass
1970 /// through the cost modeling.
1971 InstWidening getWideningDecision(Instruction *I, unsigned VF) {
1972 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1972, __PRETTY_FUNCTION__))
;
1973 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
1974 auto Itr = WideningDecisions.find(InstOnVF);
1975 if (Itr == WideningDecisions.end())
1976 return CM_Unknown;
1977 return Itr->second.first;
1978 }
1979
1980 /// Return the vectorization cost for the given instruction \p I and vector
1981 /// width \p VF.
1982 unsigned getWideningCost(Instruction *I, unsigned VF) {
1983 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1983, __PRETTY_FUNCTION__))
;
1984 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
1985 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 1985, __PRETTY_FUNCTION__))
;
1986 return WideningDecisions[InstOnVF].second;
1987 }
1988
1989 /// Return True if instruction \p I is an optimizable truncate whose operand
1990 /// is an induction variable. Such a truncate will be removed by adding a new
1991 /// induction variable with the destination type.
1992 bool isOptimizableIVTruncate(Instruction *I, unsigned VF) {
1993
1994 // If the instruction is not a truncate, return false.
1995 auto *Trunc = dyn_cast<TruncInst>(I);
1996 if (!Trunc)
1997 return false;
1998
1999 // Get the source and destination types of the truncate.
2000 Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF);
2001 Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF);
2002
2003 // If the truncate is free for the given types, return false. Replacing a
2004 // free truncate with an induction variable would add an induction variable
2005 // update instruction to each iteration of the loop. We exclude from this
2006 // check the primary induction variable since it will need an update
2007 // instruction regardless.
2008 Value *Op = Trunc->getOperand(0);
2009 if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))
2010 return false;
2011
2012 // If the truncated value is not an induction variable, return false.
2013 return Legal->isInductionVariable(Op);
2014 }
2015
2016private:
2017 /// \return An upper bound for the vectorization factor, larger than zero.
2018 /// One is returned if vectorization should best be avoided due to cost.
2019 unsigned computeFeasibleMaxVF(bool OptForSize);
2020
2021 /// The vectorization cost is a combination of the cost itself and a boolean
2022 /// indicating whether any of the contributing operations will actually
2023 /// operate on
2024 /// vector values after type legalization in the backend. If this latter value
2025 /// is
2026 /// false, then all operations will be scalarized (i.e. no vectorization has
2027 /// actually taken place).
2028 typedef std::pair<unsigned, bool> VectorizationCostTy;
2029
2030 /// Returns the expected execution cost. The unit of the cost does
2031 /// not matter because we use the 'cost' units to compare different
2032 /// vector widths. The cost that is returned is *not* normalized by
2033 /// the factor width.
2034 VectorizationCostTy expectedCost(unsigned VF);
2035
2036 /// Returns the execution time cost of an instruction for a given vector
2037 /// width. Vector width of one means scalar.
2038 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
2039
2040 /// The cost-computation logic from getInstructionCost which provides
2041 /// the vector type as an output parameter.
2042 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
2043
2044 /// Calculate vectorization cost of memory instruction \p I.
2045 unsigned getMemoryInstructionCost(Instruction *I, unsigned VF);
2046
2047 /// The cost computation for scalarized memory instruction.
2048 unsigned getMemInstScalarizationCost(Instruction *I, unsigned VF);
2049
2050 /// The cost computation for interleaving group of memory instructions.
2051 unsigned getInterleaveGroupCost(Instruction *I, unsigned VF);
2052
2053 /// The cost computation for Gather/Scatter instruction.
2054 unsigned getGatherScatterCost(Instruction *I, unsigned VF);
2055
2056 /// The cost computation for widening instruction \p I with consecutive
2057 /// memory access.
2058 unsigned getConsecutiveMemOpCost(Instruction *I, unsigned VF);
2059
2060 /// The cost calculation for Load instruction \p I with uniform pointer -
2061 /// scalar load + broadcast.
2062 unsigned getUniformMemOpCost(Instruction *I, unsigned VF);
2063
2064 /// Returns whether the instruction is a load or store and will be a emitted
2065 /// as a vector operation.
2066 bool isConsecutiveLoadOrStore(Instruction *I);
2067
2068 /// Create an analysis remark that explains why vectorization failed
2069 ///
2070 /// \p RemarkName is the identifier for the remark. \return the remark object
2071 /// that can be streamed to.
2072 OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
2073 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
2074 RemarkName, TheLoop);
2075 }
2076
2077 /// Map of scalar integer values to the smallest bitwidth they can be legally
2078 /// represented as. The vector equivalents of these values should be truncated
2079 /// to this type.
2080 MapVector<Instruction *, uint64_t> MinBWs;
2081
2082 /// A type representing the costs for instructions if they were to be
2083 /// scalarized rather than vectorized. The entries are Instruction-Cost
2084 /// pairs.
2085 typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
2086
2087 /// A set containing all BasicBlocks that are known to present after
2088 /// vectorization as a predicated block.
2089 SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization;
2090
2091 /// A map holding scalar costs for different vectorization factors. The
2092 /// presence of a cost for an instruction in the mapping indicates that the
2093 /// instruction will be scalarized when vectorizing with the associated
2094 /// vectorization factor. The entries are VF-ScalarCostTy pairs.
2095 DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
2096
2097 /// Holds the instructions known to be uniform after vectorization.
2098 /// The data is collected per VF.
2099 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Uniforms;
2100
2101 /// Holds the instructions known to be scalar after vectorization.
2102 /// The data is collected per VF.
2103 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Scalars;
2104
2105 /// Returns the expected difference in cost from scalarizing the expression
2106 /// feeding a predicated instruction \p PredInst. The instructions to
2107 /// scalarize and their scalar costs are collected in \p ScalarCosts. A
2108 /// non-negative return value implies the expression will be scalarized.
2109 /// Currently, only single-use chains are considered for scalarization.
2110 int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
2111 unsigned VF);
2112
2113 /// Collects the instructions to scalarize for each predicated instruction in
2114 /// the loop.
2115 void collectInstsToScalarize(unsigned VF);
2116
2117 /// Collect the instructions that are uniform after vectorization. An
2118 /// instruction is uniform if we represent it with a single scalar value in
2119 /// the vectorized loop corresponding to each vector iteration. Examples of
2120 /// uniform instructions include pointer operands of consecutive or
2121 /// interleaved memory accesses. Note that although uniformity implies an
2122 /// instruction will be scalar, the reverse is not true. In general, a
2123 /// scalarized instruction will be represented by VF scalar values in the
2124 /// vectorized loop, each corresponding to an iteration of the original
2125 /// scalar loop.
2126 void collectLoopUniforms(unsigned VF);
2127
2128 /// Collect the instructions that are scalar after vectorization. An
2129 /// instruction is scalar if it is known to be uniform or will be scalarized
2130 /// during vectorization. Non-uniform scalarized instructions will be
2131 /// represented by VF values in the vectorized loop, each corresponding to an
2132 /// iteration of the original scalar loop.
2133 void collectLoopScalars(unsigned VF);
2134
2135 /// Collect Uniform and Scalar values for the given \p VF.
2136 /// The sets depend on CM decision for Load/Store instructions
2137 /// that may be vectorized as interleave, gather-scatter or scalarized.
2138 void collectUniformsAndScalars(unsigned VF) {
2139 // Do the analysis once.
2140 if (VF == 1 || Uniforms.count(VF))
2141 return;
2142 setCostBasedWideningDecision(VF);
2143 collectLoopUniforms(VF);
2144 collectLoopScalars(VF);
2145 }
2146
2147 /// Keeps cost model vectorization decision and cost for instructions.
2148 /// Right now it is used for memory instructions only.
2149 typedef DenseMap<std::pair<Instruction *, unsigned>,
2150 std::pair<InstWidening, unsigned>>
2151 DecisionList;
2152
2153 DecisionList WideningDecisions;
2154
2155public:
2156 /// The loop that we evaluate.
2157 Loop *TheLoop;
2158 /// Predicated scalar evolution analysis.
2159 PredicatedScalarEvolution &PSE;
2160 /// Loop Info analysis.
2161 LoopInfo *LI;
2162 /// Vectorization legality.
2163 LoopVectorizationLegality *Legal;
2164 /// Vector target information.
2165 const TargetTransformInfo &TTI;
2166 /// Target Library Info.
2167 const TargetLibraryInfo *TLI;
2168 /// Demanded bits analysis.
2169 DemandedBits *DB;
2170 /// Assumption cache.
2171 AssumptionCache *AC;
2172 /// Interface to emit optimization remarks.
2173 OptimizationRemarkEmitter *ORE;
2174
2175 const Function *TheFunction;
2176 /// Loop Vectorize Hint.
2177 const LoopVectorizeHints *Hints;
2178 /// Values to ignore in the cost model.
2179 SmallPtrSet<const Value *, 16> ValuesToIgnore;
2180 /// Values to ignore in the cost model when VF > 1.
2181 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2182};
2183
2184/// LoopVectorizationPlanner - drives the vectorization process after having
2185/// passed Legality checks.
2186class LoopVectorizationPlanner {
2187public:
2188 LoopVectorizationPlanner(LoopVectorizationCostModel &CM) : CM(CM) {}
2189
2190 ~LoopVectorizationPlanner() {}
2191
2192 /// Plan how to best vectorize, return the best VF and its cost.
2193 LoopVectorizationCostModel::VectorizationFactor plan(bool OptForSize,
2194 unsigned UserVF);
2195
2196private:
2197 /// The profitablity analysis.
2198 LoopVectorizationCostModel &CM;
2199};
2200
2201/// \brief This holds vectorization requirements that must be verified late in
2202/// the process. The requirements are set by legalize and costmodel. Once
2203/// vectorization has been determined to be possible and profitable the
2204/// requirements can be verified by looking for metadata or compiler options.
2205/// For example, some loops require FP commutativity which is only allowed if
2206/// vectorization is explicitly specified or if the fast-math compiler option
2207/// has been provided.
2208/// Late evaluation of these requirements allows helpful diagnostics to be
2209/// composed that tells the user what need to be done to vectorize the loop. For
2210/// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2211/// evaluation should be used only when diagnostics can generated that can be
2212/// followed by a non-expert user.
2213class LoopVectorizationRequirements {
2214public:
2215 LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE)
2216 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr), ORE(ORE) {}
2217
2218 void addUnsafeAlgebraInst(Instruction *I) {
2219 // First unsafe algebra instruction.
2220 if (!UnsafeAlgebraInst)
2221 UnsafeAlgebraInst = I;
2222 }
2223
2224 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2225
2226 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2227 const char *PassName = Hints.vectorizeAnalysisPassName();
2228 bool Failed = false;
2229 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2230 ORE.emit(
2231 OptimizationRemarkAnalysisFPCommute(PassName, "CantReorderFPOps",
2232 UnsafeAlgebraInst->getDebugLoc(),
2233 UnsafeAlgebraInst->getParent())
2234 << "loop not vectorized: cannot prove it is safe to reorder "
2235 "floating-point operations");
2236 Failed = true;
2237 }
2238
2239 // Test if runtime memcheck thresholds are exceeded.
2240 bool PragmaThresholdReached =
2241 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2242 bool ThresholdReached =
2243 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2244 if ((ThresholdReached && !Hints.allowReordering()) ||
2245 PragmaThresholdReached) {
2246 ORE.emit(OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2247 L->getStartLoc(),
2248 L->getHeader())
2249 << "loop not vectorized: cannot prove it is safe to reorder "
2250 "memory operations");
2251 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)
;
2252 Failed = true;
2253 }
2254
2255 return Failed;
2256 }
2257
2258private:
2259 unsigned NumRuntimePointerChecks;
2260 Instruction *UnsafeAlgebraInst;
2261
2262 /// Interface to emit optimization remarks.
2263 OptimizationRemarkEmitter &ORE;
2264};
2265
2266static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
2267 if (L.empty()) {
2268 if (!hasCyclesInLoopBody(L))
2269 V.push_back(&L);
2270 return;
2271 }
2272 for (Loop *InnerL : L)
2273 addAcyclicInnerLoop(*InnerL, V);
2274}
2275
2276/// The LoopVectorize Pass.
2277struct LoopVectorize : public FunctionPass {
2278 /// Pass identification, replacement for typeid
2279 static char ID;
2280
2281 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2282 : FunctionPass(ID) {
2283 Impl.DisableUnrolling = NoUnrolling;
2284 Impl.AlwaysVectorize = AlwaysVectorize;
2285 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
2286 }
2287
2288 LoopVectorizePass Impl;
2289
2290 bool runOnFunction(Function &F) override {
2291 if (skipFunction(F))
2292 return false;
2293
2294 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2295 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2296 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2297 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2298 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2299 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2300 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2301 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2302 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2303 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2304 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2305 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2306
2307 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2308 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2309
2310 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2311 GetLAA, *ORE);
2312 }
2313
2314 void getAnalysisUsage(AnalysisUsage &AU) const override {
2315 AU.addRequired<AssumptionCacheTracker>();
2316 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2317 AU.addRequired<DominatorTreeWrapperPass>();
2318 AU.addRequired<LoopInfoWrapperPass>();
2319 AU.addRequired<ScalarEvolutionWrapperPass>();
2320 AU.addRequired<TargetTransformInfoWrapperPass>();
2321 AU.addRequired<AAResultsWrapperPass>();
2322 AU.addRequired<LoopAccessLegacyAnalysis>();
2323 AU.addRequired<DemandedBitsWrapperPass>();
2324 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2325 AU.addPreserved<LoopInfoWrapperPass>();
2326 AU.addPreserved<DominatorTreeWrapperPass>();
2327 AU.addPreserved<BasicAAWrapperPass>();
2328 AU.addPreserved<GlobalsAAWrapperPass>();
2329 }
2330};
2331
2332} // end anonymous namespace
2333
2334//===----------------------------------------------------------------------===//
2335// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2336// LoopVectorizationCostModel and LoopVectorizationPlanner.
2337//===----------------------------------------------------------------------===//
2338
2339Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
2340 // We need to place the broadcast of invariant variables outside the loop.
2341 Instruction *Instr = dyn_cast<Instruction>(V);
2342 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2343 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2344
2345 // Place the code for broadcasting invariant variables in the new preheader.
2346 IRBuilder<>::InsertPointGuard Guard(Builder);
2347 if (Invariant)
2348 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2349
2350 // Broadcast the scalar into all locations in the vector.
2351 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2352
2353 return Shuf;
2354}
2355
2356void InnerLoopVectorizer::createVectorIntOrFpInductionPHI(
2357 const InductionDescriptor &II, Value *Step, Instruction *EntryVal) {
2358 Value *Start = II.getStartValue();
2359
2360 // Construct the initial value of the vector IV in the vector loop preheader
2361 auto CurrIP = Builder.saveIP();
2362 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2363 if (isa<TruncInst>(EntryVal)) {
2364 assert(Start->getType()->isIntegerTy() &&((Start->getType()->isIntegerTy() && "Truncation requires an integer type"
) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2365, __PRETTY_FUNCTION__))
2365 "Truncation requires an integer type")((Start->getType()->isIntegerTy() && "Truncation requires an integer type"
) ? static_cast<void> (0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2365, __PRETTY_FUNCTION__))
;
2366 auto *TruncType = cast<IntegerType>(EntryVal->getType());
2367 Step = Builder.CreateTrunc(Step, TruncType);
2368 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2369 }
2370 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2371 Value *SteppedStart =
2372 getStepVector(SplatStart, 0, Step, II.getInductionOpcode());
2373
2374 // We create vector phi nodes for both integer and floating-point induction
2375 // variables. Here, we determine the kind of arithmetic we will perform.
2376 Instruction::BinaryOps AddOp;
2377 Instruction::BinaryOps MulOp;
2378 if (Step->getType()->isIntegerTy()) {
2379 AddOp = Instruction::Add;
2380 MulOp = Instruction::Mul;
2381 } else {
2382 AddOp = II.getInductionOpcode();
2383 MulOp = Instruction::FMul;
2384 }
2385
2386 // Multiply the vectorization factor by the step using integer or
2387 // floating-point arithmetic as appropriate.
2388 Value *ConstVF = getSignedIntOrFpConstant(Step->getType(), VF);
2389 Value *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, Step, ConstVF));
2390
2391 // Create a vector splat to use in the induction update.
2392 //
2393 // FIXME: If the step is non-constant, we create the vector splat with
2394 // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
2395 // handle a constant vector splat.
2396 Value *SplatVF = isa<Constant>(Mul)
2397 ? ConstantVector::getSplat(VF, cast<Constant>(Mul))
2398 : Builder.CreateVectorSplat(VF, Mul);
2399 Builder.restoreIP(CurrIP);
2400
2401 // We may need to add the step a number of times, depending on the unroll
2402 // factor. The last of those goes into the PHI.
2403 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2404 &*LoopVectorBody->getFirstInsertionPt());
2405 Instruction *LastInduction = VecInd;
2406 VectorParts Entry(UF);
2407 for (unsigned Part = 0; Part < UF; ++Part) {
2408 Entry[Part] = LastInduction;
2409 LastInduction = cast<Instruction>(addFastMathFlag(
2410 Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")));
2411 }
2412 VectorLoopValueMap.initVector(EntryVal, Entry);
2413 if (isa<TruncInst>(EntryVal))
2414 addMetadata(Entry, EntryVal);
2415
2416 // Move the last step to the end of the latch block. This ensures consistent
2417 // placement of all induction updates.
2418 auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2419 auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2420 auto *ICmp = cast<Instruction>(Br->getCondition());
2421 LastInduction->moveBefore(ICmp);
2422 LastInduction->setName("vec.ind.next");
2423
2424 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2425 VecInd->addIncoming(LastInduction, LoopVectorLatch);
2426}
2427
2428bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
2429 return Cost->isScalarAfterVectorization(I, VF) ||
2430 Cost->isProfitableToScalarize(I, VF);
2431}
2432
2433bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
2434 if (shouldScalarizeInstruction(IV))
2435 return true;
2436 auto isScalarInst = [&](User *U) -> bool {
2437 auto *I = cast<Instruction>(U);
2438 return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2439 };
2440 return any_of(IV->users(), isScalarInst);
2441}
2442
2443void InnerLoopVectorizer::widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc) {
2444
2445 assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&(((IV->getType()->isIntegerTy() || IV != OldInduction) &&
"Primary induction variable must have an integer type") ? static_cast
<void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2446, __PRETTY_FUNCTION__))
2446 "Primary induction variable must have an integer type")(((IV->getType()->isIntegerTy() || IV != OldInduction) &&
"Primary induction variable must have an integer type") ? static_cast
<void> (0) : __assert_fail ("(IV->getType()->isIntegerTy() || IV != OldInduction) && \"Primary induction variable must have an integer type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2446, __PRETTY_FUNCTION__))
;
2447
2448 auto II = Legal->getInductionVars()->find(IV);
2449 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2449, __PRETTY_FUNCTION__))
;
2450
2451 auto ID = II->second;
2452 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2452, __PRETTY_FUNCTION__))
;
2453
2454 // The scalar value to broadcast. This will be derived from the canonical
2455 // induction variable.
2456 Value *ScalarIV = nullptr;
2457
2458 // The value from the original loop to which we are mapping the new induction
2459 // variable.
2460 Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2461
2462 // True if we have vectorized the induction variable.
2463 auto VectorizedIV = false;
2464
2465 // Determine if we want a scalar version of the induction variable. This is
2466 // true if the induction variable itself is not widened, or if it has at
2467 // least one user in the loop that is not widened.
2468 auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2469
2470 // Generate code for the induction step. Note that induction steps are
2471 // required to be loop-invariant
2472 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2473, __PRETTY_FUNCTION__))
2473 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2473, __PRETTY_FUNCTION__))
;
2474 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2475 Value *Step = nullptr;
2476 if (PSE.getSE()->isSCEVable(IV->getType())) {
2477 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2478 Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2479 LoopVectorPreHeader->getTerminator());
2480 } else {
2481 Step = cast<SCEVUnknown>(ID.getStep())->getValue();
2482 }
2483
2484 // Try to create a new independent vector induction variable. If we can't
2485 // create the phi node, we will splat the scalar induction variable in each
2486 // loop iteration.
2487 if (VF > 1 && !shouldScalarizeInstruction(EntryVal)) {
2488 createVectorIntOrFpInductionPHI(ID, Step, EntryVal);
2489 VectorizedIV = true;
2490 }
2491
2492 // If we haven't yet vectorized the induction variable, or if we will create
2493 // a scalar one, we need to define the scalar induction variable and step
2494 // values. If we were given a truncation type, truncate the canonical
2495 // induction variable and step. Otherwise, derive these values from the
2496 // induction descriptor.
2497 if (!VectorizedIV || NeedsScalarIV) {
2498 ScalarIV = Induction;
2499 if (IV != OldInduction) {
2500 ScalarIV = IV->getType()->isIntegerTy()
2501 ? Builder.CreateSExtOrTrunc(Induction, IV->getType())
2502 : Builder.CreateCast(Instruction::SIToFP, Induction,
2503 IV->getType());
2504 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2505 ScalarIV->setName("offset.idx");
2506 }
2507 if (Trunc) {
2508 auto *TruncType = cast<IntegerType>(Trunc->getType());
2509 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2510, __PRETTY_FUNCTION__))
2510 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2510, __PRETTY_FUNCTION__))
;
2511 ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType);
2512 Step = Builder.CreateTrunc(Step, TruncType);
2513 }
2514 }
2515
2516 // If we haven't yet vectorized the induction variable, splat the scalar
2517 // induction variable, and build the necessary step vectors.
2518 if (!VectorizedIV) {
2519 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2520 VectorParts Entry(UF);
2521 for (unsigned Part = 0; Part < UF; ++Part)
2522 Entry[Part] =
2523 getStepVector(Broadcasted, VF * Part, Step, ID.getInductionOpcode());
2524 VectorLoopValueMap.initVector(EntryVal, Entry);
2525 if (Trunc)
2526 addMetadata(Entry, Trunc);
2527 }
2528
2529 // If an induction variable is only used for counting loop iterations or
2530 // calculating addresses, it doesn't need to be widened. Create scalar steps
2531 // that can be used by instructions we will later scalarize. Note that the
2532 // addition of the scalar steps will not increase the number of instructions
2533 // in the loop in the common case prior to InstCombine. We will be trading
2534 // one vector extract for each scalar step.
2535 if (NeedsScalarIV)
2536 buildScalarSteps(ScalarIV, Step, EntryVal, ID);
2537}
2538
2539Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
2540 Instruction::BinaryOps BinOp) {
2541 // Create and check the types.
2542 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2542, __PRETTY_FUNCTION__))
;
2543 int VLen = Val->getType()->getVectorNumElements();
2544
2545 Type *STy = Val->getType()->getScalarType();
2546 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2547, __PRETTY_FUNCTION__))
2547 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2547, __PRETTY_FUNCTION__))
;
2548 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2548, __PRETTY_FUNCTION__))
;
2549
2550 SmallVector<Constant *, 8> Indices;
2551
2552 if (STy->isIntegerTy()) {
2553 // Create a vector of consecutive numbers from zero to VF.
2554 for (int i = 0; i < VLen; ++i)
2555 Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2556
2557 // Add the consecutive indices to the vector value.
2558 Constant *Cv = ConstantVector::get(Indices);
2559 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2559, __PRETTY_FUNCTION__))
;
2560 Step = Builder.CreateVectorSplat(VLen, Step);
2561 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2561, __PRETTY_FUNCTION__))
;
2562 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2563 // which can be found from the original scalar operations.
2564 Step = Builder.CreateMul(Cv, Step);
2565 return Builder.CreateAdd(Val, Step, "induction");
2566 }
2567
2568 // Floating point induction.
2569 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2570, __PRETTY_FUNCTION__))
2570 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2570, __PRETTY_FUNCTION__))
;
2571 // Create a vector of consecutive numbers from zero to VF.
2572 for (int i = 0; i < VLen; ++i)
2573 Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2574
2575 // Add the consecutive indices to the vector value.
2576 Constant *Cv = ConstantVector::get(Indices);
2577
2578 Step = Builder.CreateVectorSplat(VLen, Step);
2579
2580 // Floating point operations had to be 'fast' to enable the induction.
2581 FastMathFlags Flags;
2582 Flags.setUnsafeAlgebra();
2583
2584 Value *MulOp = Builder.CreateFMul(Cv, Step);
2585 if (isa<Instruction>(MulOp))
2586 // Have to check, MulOp may be a constant
2587 cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2588
2589 Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2590 if (isa<Instruction>(BOp))
2591 cast<Instruction>(BOp)->setFastMathFlags(Flags);
2592 return BOp;
2593}
2594
2595void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2596 Value *EntryVal,
2597 const InductionDescriptor &ID) {
2598
2599 // We shouldn't have to build scalar steps if we aren't vectorizing.
2600 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2600, __PRETTY_FUNCTION__))
;
2601
2602 // Get the value type and ensure it and the step have the same integer type.
2603 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2604 assert(ScalarIVTy == Step->getType() &&((ScalarIVTy == Step->getType() && "Val and Step should have the same type"
) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2605, __PRETTY_FUNCTION__))
2605 "Val and Step should have the same type")((ScalarIVTy == Step->getType() && "Val and Step should have the same type"
) ? static_cast<void> (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2605, __PRETTY_FUNCTION__))
;
2606
2607 // We build scalar steps for both integer and floating-point induction
2608 // variables. Here, we determine the kind of arithmetic we will perform.
2609 Instruction::BinaryOps AddOp;
2610 Instruction::BinaryOps MulOp;
2611 if (ScalarIVTy->isIntegerTy()) {
2612 AddOp = Instruction::Add;
2613 MulOp = Instruction::Mul;
2614 } else {
2615 AddOp = ID.getInductionOpcode();
2616 MulOp = Instruction::FMul;
2617 }
2618
2619 // Determine the number of scalars we need to generate for each unroll
2620 // iteration. If EntryVal is uniform, we only need to generate the first
2621 // lane. Otherwise, we generate all VF values.
2622 unsigned Lanes =
2623 Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) ? 1 : VF;
2624
2625 // Compute the scalar steps and save the results in VectorLoopValueMap.
2626 ScalarParts Entry(UF);
2627 for (unsigned Part = 0; Part < UF; ++Part) {
2628 Entry[Part].resize(VF);
2629 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2630 auto *StartIdx = getSignedIntOrFpConstant(ScalarIVTy, VF * Part + Lane);
2631 auto *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, StartIdx, Step));
2632 auto *Add = addFastMathFlag(Builder.CreateBinOp(AddOp, ScalarIV, Mul));
2633 Entry[Part][Lane] = Add;
2634 }
2635 }
2636 VectorLoopValueMap.initScalar(EntryVal, Entry);
2637}
2638
2639int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2640
2641 const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2642 ValueToValueMap();
2643
2644 int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2645 if (Stride == 1 || Stride == -1)
2646 return Stride;
2647 return 0;
2648}
2649
2650bool LoopVectorizationLegality::isUniform(Value *V) {
2651 return LAI->isUniform(V);
2652}
2653
2654const InnerLoopVectorizer::VectorParts &
2655InnerLoopVectorizer::getVectorValue(Value *V) {
2656 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2656, __PRETTY_FUNCTION__))
;
2657 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2657, __PRETTY_FUNCTION__))
;
2658 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2658, __PRETTY_FUNCTION__))
;
2659
2660 // If we have a stride that is replaced by one, do it here.
2661 if (Legal->hasStride(V))
2662 V = ConstantInt::get(V->getType(), 1);
2663
2664 // If we have this scalar in the map, return it.
2665 if (VectorLoopValueMap.hasVector(V))
2666 return VectorLoopValueMap.VectorMapStorage[V];
2667
2668 // If the value has not been vectorized, check if it has been scalarized
2669 // instead. If it has been scalarized, and we actually need the value in
2670 // vector form, we will construct the vector values on demand.
2671 if (VectorLoopValueMap.hasScalar(V)) {
2672
2673 // Initialize a new vector map entry.
2674 VectorParts Entry(UF);
2675
2676 // If we've scalarized a value, that value should be an instruction.
2677 auto *I = cast<Instruction>(V);
2678
2679 // If we aren't vectorizing, we can just copy the scalar map values over to
2680 // the vector map.
2681 if (VF == 1) {
2682 for (unsigned Part = 0; Part < UF; ++Part)
2683 Entry[Part] = getScalarValue(V, Part, 0);
2684 return VectorLoopValueMap.initVector(V, Entry);
2685 }
2686
2687 // Get the last scalar instruction we generated for V. If the value is
2688 // known to be uniform after vectorization, this corresponds to lane zero
2689 // of the last unroll iteration. Otherwise, the last instruction is the one
2690 // we created for the last vector lane of the last unroll iteration.
2691 unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) ? 0 : VF - 1;
2692 auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
2693
2694 // Set the insert point after the last scalarized instruction. This ensures
2695 // the insertelement sequence will directly follow the scalar definitions.
2696 auto OldIP = Builder.saveIP();
2697 auto NewIP = std::next(BasicBlock::iterator(LastInst));
2698 Builder.SetInsertPoint(&*NewIP);
2699
2700 // However, if we are vectorizing, we need to construct the vector values.
2701 // If the value is known to be uniform after vectorization, we can just
2702 // broadcast the scalar value corresponding to lane zero for each unroll
2703 // iteration. Otherwise, we construct the vector values using insertelement
2704 // instructions. Since the resulting vectors are stored in
2705 // VectorLoopValueMap, we will only generate the insertelements once.
2706 for (unsigned Part = 0; Part < UF; ++Part) {
2707 Value *VectorValue = nullptr;
2708 if (Cost->isUniformAfterVectorization(I, VF)) {
2709 VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
2710 } else {
2711 VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
2712 for (unsigned Lane = 0; Lane < VF; ++Lane)
2713 VectorValue = Builder.CreateInsertElement(
2714 VectorValue, getScalarValue(V, Part, Lane),
2715 Builder.getInt32(Lane));
2716 }
2717 Entry[Part] = VectorValue;
2718 }
2719 Builder.restoreIP(OldIP);
2720 return VectorLoopValueMap.initVector(V, Entry);
2721 }
2722
2723 // If this scalar is unknown, assume that it is a constant or that it is
2724 // loop invariant. Broadcast V and save the value for future uses.
2725 Value *B = getBroadcastInstrs(V);
2726 return VectorLoopValueMap.initVector(V, VectorParts(UF, B));
2727}
2728
2729Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
2730 unsigned Lane) {
2731
2732 // If the value is not an instruction contained in the loop, it should
2733 // already be scalar.
2734 if (OrigLoop->isLoopInvariant(V))
2735 return V;
2736
2737 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2739, __PRETTY_FUNCTION__))
2738 !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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2739, __PRETTY_FUNCTION__))
2739 : 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2739, __PRETTY_FUNCTION__))
;
2740
2741 // If the value from the original loop has not been vectorized, it is
2742 // represented by UF x VF scalar values in the new loop. Return the requested
2743 // scalar value.
2744 if (VectorLoopValueMap.hasScalar(V))
2745 return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
2746
2747 // If the value has not been scalarized, get its entry in VectorLoopValueMap
2748 // for the given unroll part. If this entry is not a vector type (i.e., the
2749 // vectorization factor is one), there is no need to generate an
2750 // extractelement instruction.
2751 auto *U = getVectorValue(V)[Part];
2752 if (!U->getType()->isVectorTy()) {
2753 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2753, __PRETTY_FUNCTION__))
;
2754 return U;
2755 }
2756
2757 // Otherwise, the value from the original loop has been vectorized and is
2758 // represented by UF vector values. Extract and return the requested scalar
2759 // value from the appropriate vector lane.
2760 return Builder.CreateExtractElement(U, Builder.getInt32(Lane));
2761}
2762
2763Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2764 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2764, __PRETTY_FUNCTION__))
;
2765 SmallVector<Constant *, 8> ShuffleMask;
2766 for (unsigned i = 0; i < VF; ++i)
2767 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2768
2769 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2770 ConstantVector::get(ShuffleMask),
2771 "reverse");
2772}
2773
2774// Try to vectorize the interleave group that \p Instr belongs to.
2775//
2776// E.g. Translate following interleaved load group (factor = 3):
2777// for (i = 0; i < N; i+=3) {
2778// R = Pic[i]; // Member of index 0
2779// G = Pic[i+1]; // Member of index 1
2780// B = Pic[i+2]; // Member of index 2
2781// ... // do something to R, G, B
2782// }
2783// To:
2784// %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2785// %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2786// %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2787// %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2788//
2789// Or translate following interleaved store group (factor = 3):
2790// for (i = 0; i < N; i+=3) {
2791// ... do something to R, G, B
2792// Pic[i] = R; // Member of index 0
2793// Pic[i+1] = G; // Member of index 1
2794// Pic[i+2] = B; // Member of index 2
2795// }
2796// To:
2797// %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2798// %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2799// %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2800// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2801// store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2802void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2803 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2804 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2804, __PRETTY_FUNCTION__))
;
2805
2806 // Skip if current instruction is not the insert position.
2807 if (Instr != Group->getInsertPos())
2808 return;
2809
2810 Value *Ptr = getPointerOperand(Instr);
2811
2812 // Prepare for the vector type of the interleaved load/store.
2813 Type *ScalarTy = getMemInstValueType(Instr);
2814 unsigned InterleaveFactor = Group->getFactor();
2815 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2816 Type *PtrTy = VecTy->getPointerTo(getMemInstAddressSpace(Instr));
2817
2818 // Prepare for the new pointers.
2819 setDebugLocFromInst(Builder, Ptr);
2820 SmallVector<Value *, 2> NewPtrs;
2821 unsigned Index = Group->getIndex(Instr);
2822
2823 // If the group is reverse, adjust the index to refer to the last vector lane
2824 // instead of the first. We adjust the index from the first vector lane,
2825 // rather than directly getting the pointer for lane VF - 1, because the
2826 // pointer operand of the interleaved access is supposed to be uniform. For
2827 // uniform instructions, we're only required to generate a value for the
2828 // first vector lane in each unroll iteration.
2829 if (Group->isReverse())
2830 Index += (VF - 1) * Group->getFactor();
2831
2832 for (unsigned Part = 0; Part < UF; Part++) {
2833 Value *NewPtr = getScalarValue(Ptr, Part, 0);
2834
2835 // Notice current instruction could be any index. Need to adjust the address
2836 // to the member of index 0.
2837 //
2838 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2839 // b = A[i]; // Member of index 0
2840 // Current pointer is pointed to A[i+1], adjust it to A[i].
2841 //
2842 // E.g. A[i+1] = a; // Member of index 1
2843 // A[i] = b; // Member of index 0
2844 // A[i+2] = c; // Member of index 2 (Current instruction)
2845 // Current pointer is pointed to A[i+2], adjust it to A[i].
2846 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2847
2848 // Cast to the vector pointer type.
2849 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2850 }
2851
2852 setDebugLocFromInst(Builder, Instr);
2853 Value *UndefVec = UndefValue::get(VecTy);
2854
2855 // Vectorize the interleaved load group.
2856 if (isa<LoadInst>(Instr)) {
2857
2858 // For each unroll part, create a wide load for the group.
2859 SmallVector<Value *, 2> NewLoads;
2860 for (unsigned Part = 0; Part < UF; Part++) {
2861 auto *NewLoad = Builder.CreateAlignedLoad(
2862 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2863 addMetadata(NewLoad, Instr);
2864 NewLoads.push_back(NewLoad);
2865 }
2866
2867 // For each member in the group, shuffle out the appropriate data from the
2868 // wide loads.
2869 for (unsigned I = 0; I < InterleaveFactor; ++I) {
2870 Instruction *Member = Group->getMember(I);
2871
2872 // Skip the gaps in the group.
2873 if (!Member)
2874 continue;
2875
2876 VectorParts Entry(UF);
2877 Constant *StrideMask = createStrideMask(Builder, I, InterleaveFactor, VF);
2878 for (unsigned Part = 0; Part < UF; Part++) {
2879 Value *StridedVec = Builder.CreateShuffleVector(
2880 NewLoads[Part], UndefVec, StrideMask, "strided.vec");
2881
2882 // If this member has different type, cast the result type.
2883 if (Member->getType() != ScalarTy) {
2884 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2885 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2886 }
2887
2888 Entry[Part] =
2889 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2890 }
2891 VectorLoopValueMap.initVector(Member, Entry);
2892 }
2893 return;
2894 }
2895
2896 // The sub vector type for current instruction.
2897 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2898
2899 // Vectorize the interleaved store group.
2900 for (unsigned Part = 0; Part < UF; Part++) {
2901 // Collect the stored vector from each member.
2902 SmallVector<Value *, 4> StoredVecs;
2903 for (unsigned i = 0; i < InterleaveFactor; i++) {
2904 // Interleaved store group doesn't allow a gap, so each index has a member
2905 Instruction *Member = Group->getMember(i);
2906 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2906, __PRETTY_FUNCTION__))
;
2907
2908 Value *StoredVec =
2909 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2910 if (Group->isReverse())
2911 StoredVec = reverseVector(StoredVec);
2912
2913 // If this member has different type, cast it to an unified type.
2914 if (StoredVec->getType() != SubVT)
2915 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2916
2917 StoredVecs.push_back(StoredVec);
2918 }
2919
2920 // Concatenate all vectors into a wide vector.
2921 Value *WideVec = concatenateVectors(Builder, StoredVecs);
2922
2923 // Interleave the elements in the wide vector.
2924 Constant *IMask = createInterleaveMask(Builder, VF, InterleaveFactor);
2925 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2926 "interleaved.vec");
2927
2928 Instruction *NewStoreInstr =
2929 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2930 addMetadata(NewStoreInstr, Instr);
2931 }
2932}
2933
2934void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2935 // Attempt to issue a wide load.
2936 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2937 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2938
2939 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2939, __PRETTY_FUNCTION__))
;
2940
2941 LoopVectorizationCostModel::InstWidening Decision =
2942 Cost->getWideningDecision(Instr, VF);
2943 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2944, __PRETTY_FUNCTION__))
2944 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2944, __PRETTY_FUNCTION__))
;
2945 if (Decision == LoopVectorizationCostModel::CM_Interleave)
2946 return vectorizeInterleaveGroup(Instr);
2947
2948 Type *ScalarDataTy = getMemInstValueType(Instr);
2949 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2950 Value *Ptr = getPointerOperand(Instr);
2951 unsigned Alignment = getMemInstAlignment(Instr);
2952 // An alignment of 0 means target abi alignment. We need to use the scalar's
2953 // target abi alignment in such a case.
2954 const DataLayout &DL = Instr->getModule()->getDataLayout();
2955 if (!Alignment)
2956 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2957 unsigned AddressSpace = getMemInstAddressSpace(Instr);
2958
2959 // Scalarize the memory instruction if necessary.
2960 if (Decision == LoopVectorizationCostModel::CM_Scalarize)
2961 return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
2962
2963 // Determine if the pointer operand of the access is either consecutive or
2964 // reverse consecutive.
2965 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2966 bool Reverse = ConsecutiveStride < 0;
2967 bool CreateGatherScatter =
2968 (Decision == LoopVectorizationCostModel::CM_GatherScatter);
2969
2970 VectorParts VectorGep;
2971
2972 // Handle consecutive loads/stores.
2973 if (ConsecutiveStride) {
2974 Ptr = getScalarValue(Ptr, 0, 0);
2975 } else {
2976 // At this point we should vector version of GEP for Gather or Scatter
2977 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2977, __PRETTY_FUNCTION__))
;
2978 VectorGep = getVectorValue(Ptr);
2979 }
2980
2981 VectorParts Mask = createBlockInMask(Instr->getParent());
2982 // Handle Stores:
2983 if (SI) {
2984 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2985, __PRETTY_FUNCTION__))
2985 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 2985, __PRETTY_FUNCTION__))
;
2986 setDebugLocFromInst(Builder, SI);
2987 // We don't want to update the value in the map as it might be used in
2988 // another expression. So don't use a reference type for "StoredVal".
2989 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2990
2991 for (unsigned Part = 0; Part < UF; ++Part) {
2992 Instruction *NewSI = nullptr;
2993 if (CreateGatherScatter) {
2994 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
2995 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
2996 Alignment, MaskPart);
2997 } else {
2998 // Calculate the pointer for the specific unroll-part.
2999 Value *PartPtr =
3000 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3001
3002 if (Reverse) {
3003 // If we store to reverse consecutive memory locations, then we need
3004 // to reverse the order of elements in the stored value.
3005 StoredVal[Part] = reverseVector(StoredVal[Part]);
3006 // If the address is consecutive but reversed, then the
3007 // wide store needs to start at the last vector element.
3008 PartPtr =
3009 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3010 PartPtr =
3011 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3012 Mask[Part] = reverseVector(Mask[Part]);
3013 }
3014
3015 Value *VecPtr =
3016 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3017
3018 if (Legal->isMaskRequired(SI))
3019 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
3020 Mask[Part]);
3021 else
3022 NewSI =
3023 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
3024 }
3025 addMetadata(NewSI, SI);
3026 }
3027 return;
3028 }
3029
3030 // Handle loads.
3031 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3031, __PRETTY_FUNCTION__))
;
3032 setDebugLocFromInst(Builder, LI);
3033 VectorParts Entry(UF);
3034 for (unsigned Part = 0; Part < UF; ++Part) {
3035 Instruction *NewLI;
3036 if (CreateGatherScatter) {
3037 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
3038 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
3039 0, "wide.masked.gather");
3040 Entry[Part] = NewLI;
3041 } else {
3042 // Calculate the pointer for the specific unroll-part.
3043 Value *PartPtr =
3044 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3045
3046 if (Reverse) {
3047 // If the address is consecutive but reversed, then the
3048 // wide load needs to start at the last vector element.
3049 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3050 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3051 Mask[Part] = reverseVector(Mask[Part]);
3052 }
3053
3054 Value *VecPtr =
3055 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3056 if (Legal->isMaskRequired(LI))
3057 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
3058 UndefValue::get(DataTy),
3059 "wide.masked.load");
3060 else
3061 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
3062 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
3063 }
3064 addMetadata(NewLI, LI);
3065 }
3066 VectorLoopValueMap.initVector(Instr, Entry);
3067}
3068
3069void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
3070 bool IfPredicateInstr) {
3071 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3071, __PRETTY_FUNCTION__))
;
3072 DEBUG(dbgs() << "LV: Scalarizing"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
3073 << (IfPredicateInstr ? " and predicating:" : ":") << *Instrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
3074 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Scalarizing" <<
(IfPredicateInstr ? " and predicating:" : ":") << *Instr
<< '\n'; } } while (false)
;
3075 // Holds vector parameters or scalars, in case of uniform vals.
3076 SmallVector<VectorParts, 4> Params;
3077
3078 setDebugLocFromInst(Builder, Instr);
3079
3080 // Does this instruction return a value ?
3081 bool IsVoidRetTy = Instr->getType()->isVoidTy();
3082
3083 // Initialize a new scalar map entry.
3084 ScalarParts Entry(UF);
3085
3086 VectorParts Cond;
3087 if (IfPredicateInstr)
3088 Cond = createBlockInMask(Instr->getParent());
3089
3090 // Determine the number of scalars we need to generate for each unroll
3091 // iteration. If the instruction is uniform, we only need to generate the
3092 // first lane. Otherwise, we generate all VF values.
3093 unsigned Lanes = Cost->isUniformAfterVectorization(Instr, VF) ? 1 : VF;
3094
3095 // For each vector unroll 'part':
3096 for (unsigned Part = 0; Part < UF; ++Part) {
3097 Entry[Part].resize(VF);
3098 // For each scalar that we create:
3099 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
3100
3101 // Start if-block.
3102 Value *Cmp = nullptr;
3103 if (IfPredicateInstr) {
3104 Cmp = Cond[Part];
3105 if (Cmp->getType()->isVectorTy())
3106 Cmp = Builder.CreateExtractElement(Cmp, Builder.getInt32(Lane));
3107 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
3108 ConstantInt::get(Cmp->getType(), 1));
3109 }
3110
3111 Instruction *Cloned = Instr->clone();
3112 if (!IsVoidRetTy)
3113 Cloned->setName(Instr->getName() + ".cloned");
3114
3115 // Replace the operands of the cloned instructions with their scalar
3116 // equivalents in the new loop.
3117 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3118 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, Lane);
3119 Cloned->setOperand(op, NewOp);
3120 }
3121 addNewMetadata(Cloned, Instr);
3122
3123 // Place the cloned scalar in the new loop.
3124 Builder.Insert(Cloned);
3125
3126 // Add the cloned scalar to the scalar map entry.
3127 Entry[Part][Lane] = Cloned;
3128
3129 // If we just cloned a new assumption, add it the assumption cache.
3130 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3131 if (II->getIntrinsicID() == Intrinsic::assume)
3132 AC->registerAssumption(II);
3133
3134 // End if-block.
3135 if (IfPredicateInstr)
3136 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
3137 }
3138 }
3139 VectorLoopValueMap.initScalar(Instr, Entry);
3140}
3141
3142PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
3143 Value *End, Value *Step,
3144 Instruction *DL) {
3145 BasicBlock *Header = L->getHeader();
3146 BasicBlock *Latch = L->getLoopLatch();
3147 // As we're just creating this loop, it's possible no latch exists
3148 // yet. If so, use the header as this will be a single block loop.
3149 if (!Latch)
3150 Latch = Header;
3151
3152 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
3153 Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction);
3154 setDebugLocFromInst(Builder, OldInst);
3155 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3156
3157 Builder.SetInsertPoint(Latch->getTerminator());
3158 setDebugLocFromInst(Builder, OldInst);
3159
3160 // Create i+1 and fill the PHINode.
3161 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3162 Induction->addIncoming(Start, L->getLoopPreheader());
3163 Induction->addIncoming(Next, Latch);
3164 // Create the compare.
3165 Value *ICmp = Builder.CreateICmpEQ(Next, End);
3166 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3167
3168 // Now we have two terminators. Remove the old one from the block.
3169 Latch->getTerminator()->eraseFromParent();
3170
3171 return Induction;
3172}
3173
3174Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
3175 if (TripCount)
3176 return TripCount;
3177
3178 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3179 // Find the loop boundaries.
3180 ScalarEvolution *SE = PSE.getSE();
3181 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3182 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3183, __PRETTY_FUNCTION__))
3183 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3183, __PRETTY_FUNCTION__))
;
3184
3185 Type *IdxTy = Legal->getWidestInductionType();
3186
3187 // The exit count might have the type of i64 while the phi is i32. This can
3188 // happen if we have an induction variable that is sign extended before the
3189 // compare. The only way that we get a backedge taken count is that the
3190 // induction variable was signed and as such will not overflow. In such a case
3191 // truncation is legal.
3192 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3193 IdxTy->getPrimitiveSizeInBits())
3194 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3195 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3196
3197 // Get the total trip count from the count by adding 1.
3198 const SCEV *ExitCount = SE->getAddExpr(
3199 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3200
3201 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3202
3203 // Expand the trip count and place the new instructions in the preheader.
3204 // Notice that the pre-header does not change, only the loop body.
3205 SCEVExpander Exp(*SE, DL, "induction");
3206
3207 // Count holds the overall loop count (N).
3208 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3209 L->getLoopPreheader()->getTerminator());
3210
3211 if (TripCount->getType()->isPointerTy())
3212 TripCount =
3213 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3214 L->getLoopPreheader()->getTerminator());
3215
3216 return TripCount;
3217}
3218
3219Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
3220 if (VectorTripCount)
3221 return VectorTripCount;
3222
3223 Value *TC = getOrCreateTripCount(L);
3224 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3225
3226 // Now we need to generate the expression for the part of the loop that the
3227 // vectorized body will execute. This is equal to N - (N % Step) if scalar
3228 // iterations are not required for correctness, or N - Step, otherwise. Step
3229 // is equal to the vectorization factor (number of SIMD elements) times the
3230 // unroll factor (number of SIMD instructions).
3231 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3232 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3233
3234 // If there is a non-reversed interleaved group that may speculatively access
3235 // memory out-of-bounds, we need to ensure that there will be at least one
3236 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3237 // the trip count, we set the remainder to be equal to the step. If the step
3238 // does not evenly divide the trip count, no adjustment is necessary since
3239 // there will already be scalar iterations. Note that the minimum iterations
3240 // check ensures that N >= Step.
3241 if (VF > 1 && Legal->requiresScalarEpilogue()) {
3242 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3243 R = Builder.CreateSelect(IsZero, Step, R);
3244 }
3245
3246 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3247
3248 return VectorTripCount;
3249}
3250
3251void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
3252 BasicBlock *Bypass) {
3253 Value *Count = getOrCreateTripCount(L);
3254 BasicBlock *BB = L->getLoopPreheader();
3255 IRBuilder<> Builder(BB->getTerminator());
3256
3257 // Generate code to check that the loop's trip count that we computed by
3258 // adding one to the backedge-taken count will not overflow.
3259 Value *CheckMinIters = Builder.CreateICmpULT(
3260 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3261
3262 BasicBlock *NewBB =
3263 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
3264 // Update dominator tree immediately if the generated block is a
3265 // LoopBypassBlock because SCEV expansions to generate loop bypass
3266 // checks may query it before the current function is finished.
3267 DT->addNewBlock(NewBB, BB);
3268 if (L->getParentLoop())
3269 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3270 ReplaceInstWithInst(BB->getTerminator(),
3271 BranchInst::Create(Bypass, NewBB, CheckMinIters));
3272 LoopBypassBlocks.push_back(BB);
3273}
3274
3275void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
3276 BasicBlock *Bypass) {
3277 Value *TC = getOrCreateVectorTripCount(L);
3278 BasicBlock *BB = L->getLoopPreheader();
3279 IRBuilder<> Builder(BB->getTerminator());
3280
3281 // Now, compare the new count to zero. If it is zero skip the vector loop and
3282 // jump to the scalar loop.
3283 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
3284 "cmp.zero");
3285
3286 // Generate code to check that the loop's trip count that we computed by
3287 // adding one to the backedge-taken count will not overflow.
3288 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3289 // Update dominator tree immediately if the generated block is a
3290 // LoopBypassBlock because SCEV expansions to generate loop bypass
3291 // checks may query it before the current function is finished.
3292 DT->addNewBlock(NewBB, BB);
3293 if (L->getParentLoop())
3294 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3295 ReplaceInstWithInst(BB->getTerminator(),
3296 BranchInst::Create(Bypass, NewBB, Cmp));
3297 LoopBypassBlocks.push_back(BB);
3298}
3299
3300void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
3301 BasicBlock *BB = L->getLoopPreheader();
3302
3303 // Generate the code to check that the SCEV assumptions that we made.
3304 // We want the new basic block to start at the first instruction in a
3305 // sequence of instructions that form a check.
3306 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3307 "scev.check");
3308 Value *SCEVCheck =
3309 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3310
3311 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3312 if (C->isZero())
3313 return;
3314
3315 // Create a new block containing the stride check.
3316 BB->setName("vector.scevcheck");
3317 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3318 // Update dominator tree immediately if the generated block is a
3319 // LoopBypassBlock because SCEV expansions to generate loop bypass
3320 // checks may query it before the current function is finished.
3321 DT->addNewBlock(NewBB, BB);
3322 if (L->getParentLoop())
3323 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3324 ReplaceInstWithInst(BB->getTerminator(),
3325 BranchInst::Create(Bypass, NewBB, SCEVCheck));
3326 LoopBypassBlocks.push_back(BB);
3327 AddedSafetyChecks = true;
3328}
3329
3330void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
3331 BasicBlock *BB = L->getLoopPreheader();
3332
3333 // Generate the code that checks in runtime if arrays overlap. We put the
3334 // checks into a separate block to make the more common case of few elements
3335 // faster.
3336 Instruction *FirstCheckInst;
3337 Instruction *MemRuntimeCheck;
3338 std::tie(FirstCheckInst, MemRuntimeCheck) =
3339 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3340 if (!MemRuntimeCheck)
3341 return;
3342
3343 // Create a new block containing the memory check.
3344 BB->setName("vector.memcheck");
3345 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3346 // Update dominator tree immediately if the generated block is a
3347 // LoopBypassBlock because SCEV expansions to generate loop bypass
3348 // checks may query it before the current function is finished.
3349 DT->addNewBlock(NewBB, BB);
3350 if (L->getParentLoop())
3351 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3352 ReplaceInstWithInst(BB->getTerminator(),
3353 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3354 LoopBypassBlocks.push_back(BB);
3355 AddedSafetyChecks = true;
3356
3357 // We currently don't use LoopVersioning for the actual loop cloning but we
3358 // still use it to add the noalias metadata.
3359 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3360 PSE.getSE());
3361 LVer->prepareNoAliasMetadata();
3362}
3363
3364void InnerLoopVectorizer::createEmptyLoop() {
3365 /*
3366 In this function we generate a new loop. The new loop will contain
3367 the vectorized instructions while the old loop will continue to run the
3368 scalar remainder.
3369
3370 [ ] <-- loop iteration number check.
3371 / |
3372 / v
3373 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3374 | / |
3375 | / v
3376 || [ ] <-- vector pre header.
3377 |/ |
3378 | v
3379 | [ ] \
3380 | [ ]_| <-- vector loop.
3381 | |
3382 | v
3383 | -[ ] <--- middle-block.
3384 | / |
3385 | / v
3386 -|- >[ ] <--- new preheader.
3387 | |
3388 | v
3389 | [ ] \
3390 | [ ]_| <-- old scalar loop to handle remainder.
3391 \ |
3392 \ v
3393 >[ ] <-- exit block.
3394 ...
3395 */
3396
3397 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3398 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3399 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3400 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3400, __PRETTY_FUNCTION__))
;
3401 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3401, __PRETTY_FUNCTION__))
;
3402
3403 // Some loops have a single integer induction variable, while other loops
3404 // don't. One example is c++ iterators that often have multiple pointer
3405 // induction variables. In the code below we also support a case where we
3406 // don't have a single induction variable.
3407 //
3408 // We try to obtain an induction variable from the original loop as hard
3409 // as possible. However if we don't find one that:
3410 // - is an integer
3411 // - counts from zero, stepping by one
3412 // - is the size of the widest induction variable type
3413 // then we create a new one.
3414 OldInduction = Legal->getPrimaryInduction();
3415 Type *IdxTy = Legal->getWidestInductionType();
3416
3417 // Split the single block loop into the two loop structure described above.
3418 BasicBlock *VecBody =
3419 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3420 BasicBlock *MiddleBlock =
3421 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3422 BasicBlock *ScalarPH =
3423 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3424
3425 // Create and register the new vector loop.
3426 Loop *Lp = new Loop();
3427 Loop *ParentLoop = OrigLoop->getParentLoop();
3428
3429 // Insert the new loop into the loop nest and register the new basic blocks
3430 // before calling any utilities such as SCEV that require valid LoopInfo.
3431 if (ParentLoop) {
3432 ParentLoop->addChildLoop(Lp);
3433 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3434 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3435 } else {
3436 LI->addTopLevelLoop(Lp);
3437 }
3438 Lp->addBasicBlockToLoop(VecBody, *LI);
3439
3440 // Find the loop boundaries.
3441 Value *Count = getOrCreateTripCount(Lp);
3442
3443 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3444
3445 // We need to test whether the backedge-taken count is uint##_max. Adding one
3446 // to it will cause overflow and an incorrect loop trip count in the vector
3447 // body. In case of overflow we want to directly jump to the scalar remainder
3448 // loop.
3449 emitMinimumIterationCountCheck(Lp, ScalarPH);
3450 // Now, compare the new count to zero. If it is zero skip the vector loop and
3451 // jump to the scalar loop.
3452 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3453 // Generate the code to check any assumptions that we've made for SCEV
3454 // expressions.
3455 emitSCEVChecks(Lp, ScalarPH);
3456
3457 // Generate the code that checks in runtime if arrays overlap. We put the
3458 // checks into a separate block to make the more common case of few elements
3459 // faster.
3460 emitMemRuntimeChecks(Lp, ScalarPH);
3461
3462 // Generate the induction variable.
3463 // The loop step is equal to the vectorization factor (num of SIMD elements)
3464 // times the unroll factor (num of SIMD instructions).
3465 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3466 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3467 Induction =
3468 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3469 getDebugLocFromInstOrOperands(OldInduction));
3470
3471 // We are going to resume the execution of the scalar loop.
3472 // Go over all of the induction variables that we found and fix the
3473 // PHIs that are left in the scalar version of the loop.
3474 // The starting values of PHI nodes depend on the counter of the last
3475 // iteration in the vectorized loop.
3476 // If we come from a bypass edge then we need to start from the original
3477 // start value.
3478
3479 // This variable saves the new starting index for the scalar loop. It is used
3480 // to test if there are any tail iterations left once the vector loop has
3481 // completed.
3482 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3483 for (auto &InductionEntry : *List) {
3484 PHINode *OrigPhi = InductionEntry.first;
3485 InductionDescriptor II = InductionEntry.second;
3486
3487 // Create phi nodes to merge from the backedge-taken check block.
3488 PHINode *BCResumeVal = PHINode::Create(
3489 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3490 Value *&EndValue = IVEndValues[OrigPhi];
3491 if (OrigPhi == OldInduction) {
3492 // We know what the end value is.
3493 EndValue = CountRoundDown;
3494 } else {
3495 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3496 Type *StepType = II.getStep()->getType();
3497 Instruction::CastOps CastOp =
3498 CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3499 Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3500 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3501 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3502 EndValue->setName("ind.end");
3503 }
3504
3505 // The new PHI merges the original incoming value, in case of a bypass,
3506 // or the value at the end of the vectorized loop.
3507 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3508
3509 // Fix the scalar body counter (PHI node).
3510 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3511
3512 // The old induction's phi node in the scalar body needs the truncated
3513 // value.
3514 for (BasicBlock *BB : LoopBypassBlocks)
3515 BCResumeVal->addIncoming(II.getStartValue(), BB);
3516 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3517 }
3518
3519 // Add a check in the middle block to see if we have completed
3520 // all of the iterations in the first vector loop.
3521 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3522 Value *CmpN =
3523 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3524 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3525 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3526 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3527
3528 // Get ready to start creating new instructions into the vectorized body.
3529 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3530
3531 // Save the state.
3532 LoopVectorPreHeader = Lp->getLoopPreheader();
3533 LoopScalarPreHeader = ScalarPH;
3534 LoopMiddleBlock = MiddleBlock;
3535 LoopExitBlock = ExitBlock;
3536 LoopVectorBody = VecBody;
3537 LoopScalarBody = OldBasicBlock;
3538
3539 // Keep all loop hints from the original loop on the vector loop (we'll
3540 // replace the vectorizer-specific hints below).
3541 if (MDNode *LID = OrigLoop->getLoopID())
3542 Lp->setLoopID(LID);
3543
3544 LoopVectorizeHints Hints(Lp, true, *ORE);
3545 Hints.setAlreadyVectorized();
3546}
3547
3548// Fix up external users of the induction variable. At this point, we are
3549// in LCSSA form, with all external PHIs that use the IV having one input value,
3550// coming from the remainder loop. We need those PHIs to also have a correct
3551// value for the IV when arriving directly from the middle block.
3552void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3553 const InductionDescriptor &II,
3554 Value *CountRoundDown, Value *EndValue,
3555 BasicBlock *MiddleBlock) {
3556 // There are two kinds of external IV usages - those that use the value
3557 // computed in the last iteration (the PHI) and those that use the penultimate
3558 // value (the value that feeds into the phi from the loop latch).
3559 // We allow both, but they, obviously, have different values.
3560
3561 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3561, __PRETTY_FUNCTION__))
;
3562
3563 DenseMap<Value *, Value *> MissingVals;
3564
3565 // An external user of the last iteration's value should see the value that
3566 // the remainder loop uses to initialize its own IV.
3567 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3568 for (User *U : PostInc->users()) {
3569 Instruction *UI = cast<Instruction>(U);
3570 if (!OrigLoop->contains(UI)) {
3571 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3571, __PRETTY_FUNCTION__))
;
3572 MissingVals[UI] = EndValue;
3573 }
3574 }
3575
3576 // An external user of the penultimate value need to see EndValue - Step.
3577 // The simplest way to get this is to recompute it from the constituent SCEVs,
3578 // that is Start + (Step * (CRD - 1)).
3579 for (User *U : OrigPhi->users()) {
3580 auto *UI = cast<Instruction>(U);
3581 if (!OrigLoop->contains(UI)) {
3582 const DataLayout &DL =
3583 OrigLoop->getHeader()->getModule()->getDataLayout();
3584 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3584, __PRETTY_FUNCTION__))
;
3585
3586 IRBuilder<> B(MiddleBlock->getTerminator());
3587 Value *CountMinusOne = B.CreateSub(
3588 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3589 Value *CMO =
3590 !II.getStep()->getType()->isIntegerTy()
3591 ? B.CreateCast(Instruction::SIToFP, CountMinusOne,
3592 II.getStep()->getType())
3593 : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType());
3594 CMO->setName("cast.cmo");
3595 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3596 Escape->setName("ind.escape");
3597 MissingVals[UI] = Escape;
3598 }
3599 }
3600
3601 for (auto &I : MissingVals) {
3602 PHINode *PHI = cast<PHINode>(I.first);
3603 // One corner case we have to handle is two IVs "chasing" each-other,
3604 // that is %IV2 = phi [...], [ %IV1, %latch ]
3605 // In this case, if IV1 has an external use, we need to avoid adding both
3606 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3607 // don't already have an incoming value for the middle block.
3608 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3609 PHI->addIncoming(I.second, MiddleBlock);
3610 }
3611}
3612
3613namespace {
3614struct CSEDenseMapInfo {
3615 static bool canHandle(const Instruction *I) {
3616 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3617 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3618 }
3619 static inline Instruction *getEmptyKey() {
3620 return DenseMapInfo<Instruction *>::getEmptyKey();
3621 }
3622 static inline Instruction *getTombstoneKey() {
3623 return DenseMapInfo<Instruction *>::getTombstoneKey();
3624 }
3625 static unsigned getHashValue(const Instruction *I) {
3626 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3626, __PRETTY_FUNCTION__))
;
3627 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3628 I->value_op_end()));
3629 }
3630 static bool isEqual(const Instruction *LHS, const Instruction *RHS) {
3631 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3632 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3633 return LHS == RHS;
3634 return LHS->isIdenticalTo(RHS);
3635 }
3636};
3637}
3638
3639///\brief Perform cse of induction variable instructions.
3640static void cse(BasicBlock *BB) {
3641 // Perform simple cse.
3642 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3643 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3644 Instruction *In = &*I++;
3645
3646 if (!CSEDenseMapInfo::canHandle(In))
3647 continue;
3648
3649 // Check if we can replace this instruction with any of the
3650 // visited instructions.
3651 if (Instruction *V = CSEMap.lookup(In)) {
3652 In->replaceAllUsesWith(V);
3653 In->eraseFromParent();
3654 continue;
3655 }
3656
3657 CSEMap[In] = In;
3658 }
3659}
3660
3661/// \brief Estimate the overhead of scalarizing an instruction. This is a
3662/// convenience wrapper for the type-based getScalarizationOverhead API.
3663static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3664 const TargetTransformInfo &TTI) {
3665 if (VF == 1)
3666 return 0;
3667
3668 unsigned Cost = 0;
3669 Type *RetTy = ToVectorTy(I->getType(), VF);
3670 if (!RetTy->isVoidTy() &&
3671 (!isa<LoadInst>(I) ||
3672 !TTI.supportsEfficientVectorElementLoadStore()))
3673 Cost += TTI.getScalarizationOverhead(RetTy, true, false);
3674
3675 if (CallInst *CI = dyn_cast<CallInst>(I)) {
3676 SmallVector<const Value *, 4> Operands(CI->arg_operands());
3677 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3678 }
3679 else if (!isa<StoreInst>(I) ||
3680 !TTI.supportsEfficientVectorElementLoadStore()) {
3681 SmallVector<const Value *, 4> Operands(I->operand_values());
3682 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3683 }
3684
3685 return Cost;
3686}
3687
3688// Estimate cost of a call instruction CI if it were vectorized with factor VF.
3689// Return the cost of the instruction, including scalarization overhead if it's
3690// needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3691// i.e. either vector version isn't available, or is too expensive.
3692static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3693 const TargetTransformInfo &TTI,
3694 const TargetLibraryInfo *TLI,
3695 bool &NeedToScalarize) {
3696 Function *F = CI->getCalledFunction();
3697 StringRef FnName = CI->getCalledFunction()->getName();
3698 Type *ScalarRetTy = CI->getType();
3699 SmallVector<Type *, 4> Tys, ScalarTys;
3700 for (auto &ArgOp : CI->arg_operands())
3701 ScalarTys.push_back(ArgOp->getType());
3702
3703 // Estimate cost of scalarized vector call. The source operands are assumed
3704 // to be vectors, so we need to extract individual elements from there,
3705 // execute VF scalar calls, and then gather the result into the vector return
3706 // value.
3707 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3708 if (VF == 1)
3709 return ScalarCallCost;
3710
3711 // Compute corresponding vector type for return value and arguments.
3712 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3713 for (Type *ScalarTy : ScalarTys)
3714 Tys.push_back(ToVectorTy(ScalarTy, VF));
3715
3716 // Compute costs of unpacking argument values for the scalar calls and
3717 // packing the return values to a vector.
3718 unsigned ScalarizationCost = getScalarizationOverhead(CI, VF, TTI);
3719
3720 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3721
3722 // If we can't emit a vector call for this function, then the currently found
3723 // cost is the cost we need to return.
3724 NeedToScalarize = true;
3725 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3726 return Cost;
3727
3728 // If the corresponding vector cost is cheaper, return its cost.
3729 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3730 if (VectorCallCost < Cost) {
3731 NeedToScalarize = false;
3732 return VectorCallCost;
3733 }
3734 return Cost;
3735}
3736
3737// Estimate cost of an intrinsic call instruction CI if it were vectorized with
3738// factor VF. Return the cost of the instruction, including scalarization
3739// overhead if it's needed.
3740static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3741 const TargetTransformInfo &TTI,
3742 const TargetLibraryInfo *TLI) {
3743 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3744 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3744, __PRETTY_FUNCTION__))
;
3745
3746 FastMathFlags FMF;
3747 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3748 FMF = FPMO->getFastMathFlags();
3749
3750 SmallVector<Value *, 4> Operands(CI->arg_operands());
3751 return TTI.getIntrinsicInstrCost(ID, CI->getType(), Operands, FMF, VF);
3752}
3753
3754static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3755 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3756 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3757 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3758}
3759static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3760 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3761 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3762 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3763}
3764
3765void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3766 // For every instruction `I` in MinBWs, truncate the operands, create a
3767 // truncated version of `I` and reextend its result. InstCombine runs
3768 // later and will remove any ext/trunc pairs.
3769 //
3770 SmallPtrSet<Value *, 4> Erased;
3771 for (const auto &KV : Cost->getMinimalBitwidths()) {
3772 // If the value wasn't vectorized, we must maintain the original scalar
3773 // type. The absence of the value from VectorLoopValueMap indicates that it
3774 // wasn't vectorized.
3775 if (!VectorLoopValueMap.hasVector(KV.first))
3776 continue;
3777 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3778 for (Value *&I : Parts) {
3779 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3780 continue;
3781 Type *OriginalTy = I->getType();
3782 Type *ScalarTruncatedTy =
3783 IntegerType::get(OriginalTy->getContext(), KV.second);
3784 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3785 OriginalTy->getVectorNumElements());
3786 if (TruncatedTy == OriginalTy)
3787 continue;
3788
3789 IRBuilder<> B(cast<Instruction>(I));
3790 auto ShrinkOperand = [&](Value *V) -> Value * {
3791 if (auto *ZI = dyn_cast<ZExtInst>(V))
3792 if (ZI->getSrcTy() == TruncatedTy)
3793 return ZI->getOperand(0);
3794 return B.CreateZExtOrTrunc(V, TruncatedTy);
3795 };
3796
3797 // The actual instruction modification depends on the instruction type,
3798 // unfortunately.
3799 Value *NewI = nullptr;
3800 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3801 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3802 ShrinkOperand(BO->getOperand(1)));
3803 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3804 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3805 NewI =
3806 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3807 ShrinkOperand(CI->getOperand(1)));
3808 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3809 NewI = B.CreateSelect(SI->getCondition(),
3810 ShrinkOperand(SI->getTrueValue()),
3811 ShrinkOperand(SI->getFalseValue()));
3812 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3813 switch (CI->getOpcode()) {
3814 default:
3815 llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3815)
;
3816 case Instruction::Trunc:
3817 NewI = ShrinkOperand(CI->getOperand(0));
3818 break;
3819 case Instruction::SExt:
3820 NewI = B.CreateSExtOrTrunc(
3821 CI->getOperand(0),
3822 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3823 break;
3824 case Instruction::ZExt:
3825 NewI = B.CreateZExtOrTrunc(
3826 CI->getOperand(0),
3827 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3828 break;
3829 }
3830 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3831 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3832 auto *O0 = B.CreateZExtOrTrunc(
3833 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3834 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3835 auto *O1 = B.CreateZExtOrTrunc(
3836 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3837
3838 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3839 } else if (isa<LoadInst>(I)) {
3840 // Don't do anything with the operands, just extend the result.
3841 continue;
3842 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3843 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3844 auto *O0 = B.CreateZExtOrTrunc(
3845 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3846 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3847 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3848 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3849 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3850 auto *O0 = B.CreateZExtOrTrunc(
3851 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3852 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3853 } else {
3854 llvm_unreachable("Unhandled instruction type!")::llvm::llvm_unreachable_internal("Unhandled instruction type!"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 3854)
;
3855 }
3856
3857 // Lastly, extend the result.
3858 NewI->takeName(cast<Instruction>(I));
3859 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3860 I->replaceAllUsesWith(Res);
3861 cast<Instruction>(I)->eraseFromParent();
3862 Erased.insert(I);
3863 I = Res;
3864 }
3865 }
3866
3867 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3868 for (const auto &KV : Cost->getMinimalBitwidths()) {
3869 // If the value wasn't vectorized, we must maintain the original scalar
3870 // type. The absence of the value from VectorLoopValueMap indicates that it
3871 // wasn't vectorized.
3872 if (!VectorLoopValueMap.hasVector(KV.first))
3873 continue;
3874 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3875 for (Value *&I : Parts) {
3876 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3877 if (Inst && Inst->use_empty()) {
3878 Value *NewI = Inst->getOperand(0);
3879 Inst->eraseFromParent();
3880 I = NewI;
3881 }
3882 }
3883 }
3884}
3885
3886void InnerLoopVectorizer::vectorizeLoop() {
3887 //===------------------------------------------------===//
3888 //
3889 // Notice: any optimization or new instruction that go
3890 // into the code below should be also be implemented in
3891 // the cost-model.
3892 //
3893 //===------------------------------------------------===//
3894
3895 // Collect instructions from the original loop that will become trivially dead
3896 // in the vectorized loop. We don't need to vectorize these instructions. For
3897 // example, original induction update instructions can become dead because we
3898 // separately emit induction "steps" when generating code for the new loop.
3899 // Similarly, we create a new latch condition when setting up the structure
3900 // of the new loop, so the old one can become dead.
3901 SmallPtrSet<Instruction *, 4> DeadInstructions;
3902 collectTriviallyDeadInstructions(DeadInstructions);
3903
3904 // Scan the loop in a topological order to ensure that defs are vectorized
3905 // before users.
3906 LoopBlocksDFS DFS(OrigLoop);
3907 DFS.perform(LI);
3908
3909 // Vectorize all instructions in the original loop that will not become
3910 // trivially dead when vectorized.
3911 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
3912 for (Instruction &I : *BB)
3913 if (!DeadInstructions.count(&I))
3914 vectorizeInstruction(I);
3915
3916 // Insert truncates and extends for any truncated instructions as hints to
3917 // InstCombine.
3918 if (VF > 1)
3919 truncateToMinimalBitwidths();
3920
3921 // At this point every instruction in the original loop is widened to a
3922 // vector form. Now we need to fix the recurrences in the loop. These PHI
3923 // nodes are currently empty because we did not want to introduce cycles.
3924 // This is the second stage of vectorizing recurrences.
3925 fixCrossIterationPHIs();
3926
3927 // Update the dominator tree.
3928 //
3929 // FIXME: After creating the structure of the new loop, the dominator tree is
3930 // no longer up-to-date, and it remains that way until we update it
3931 // here. An out-of-date dominator tree is problematic for SCEV,
3932 // because SCEVExpander uses it to guide code generation. The
3933 // vectorizer use SCEVExpanders in several places. Instead, we should
3934 // keep the dominator tree up-to-date as we go.
3935 updateAnalysis();
3936
3937 // Fix-up external users of the induction variables.
3938 for (auto &Entry : *Legal->getInductionVars())
3939 fixupIVUsers(Entry.first, Entry.second,
3940 getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)),
3941 IVEndValues[Entry.first], LoopMiddleBlock);
3942
3943 fixLCSSAPHIs();
3944 predicateInstructions();
3945
3946 // Remove redundant induction instructions.
3947 cse(LoopVectorBody);
3948}
3949
3950void InnerLoopVectorizer::fixCrossIterationPHIs() {
3951 // In order to support recurrences we need to be able to vectorize Phi nodes.
3952 // Phi nodes have cycles, so we need to vectorize them in two stages. This is
3953 // stage #2: We now need to fix the recurrences by adding incoming edges to
3954 // the currently empty PHI nodes. At this point every instruction in the
3955 // original loop is widened to a vector form so we can use them to construct
3956 // the incoming edges.
3957 for (Instruction &I : *OrigLoop->getHeader()) {
3958 PHINode *Phi = dyn_cast<PHINode>(&I);
3959 if (!Phi)
3960 break;
3961 // Handle first-order recurrences and reductions that need to be fixed.
3962 if (Legal->isFirstOrderRecurrence(Phi))
3963 fixFirstOrderRecurrence(Phi);
3964 else if (Legal->isReductionVariable(Phi))
3965 fixReduction(Phi);
3966 }
3967}
3968
3969void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
3970
3971 // This is the second phase of vectorizing first-order recurrences. An
3972 // overview of the transformation is described below. Suppose we have the
3973 // following loop.
3974 //
3975 // for (int i = 0; i < n; ++i)
3976 // b[i] = a[i] - a[i - 1];
3977 //
3978 // There is a first-order recurrence on "a". For this loop, the shorthand
3979 // scalar IR looks like:
3980 //
3981 // scalar.ph:
3982 // s_init = a[-1]
3983 // br scalar.body
3984 //
3985 // scalar.body:
3986 // i = phi [0, scalar.ph], [i+1, scalar.body]
3987 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
3988 // s2 = a[i]
3989 // b[i] = s2 - s1
3990 // br cond, scalar.body, ...
3991 //
3992 // In this example, s1 is a recurrence because it's value depends on the
3993 // previous iteration. In the first phase of vectorization, we created a
3994 // temporary value for s1. We now complete the vectorization and produce the
3995 // shorthand vector IR shown below (for VF = 4, UF = 1).
3996 //
3997 // vector.ph:
3998 // v_init = vector(..., ..., ..., a[-1])
3999 // br vector.body
4000 //
4001 // vector.body
4002 // i = phi [0, vector.ph], [i+4, vector.body]
4003 // v1 = phi [v_init, vector.ph], [v2, vector.body]
4004 // v2 = a[i, i+1, i+2, i+3];
4005 // v3 = vector(v1(3), v2(0, 1, 2))
4006 // b[i, i+1, i+2, i+3] = v2 - v3
4007 // br cond, vector.body, middle.block
4008 //
4009 // middle.block:
4010 // x = v2(3)
4011 // br scalar.ph
4012 //
4013 // scalar.ph:
4014 // s_init = phi [x, middle.block], [a[-1], otherwise]
4015 // br scalar.body
4016 //
4017 // After execution completes the vector loop, we extract the next value of
4018 // the recurrence (x) to use as the initial value in the scalar loop.
4019
4020 // Get the original loop preheader and single loop latch.
4021 auto *Preheader = OrigLoop->getLoopPreheader();
4022 auto *Latch = OrigLoop->getLoopLatch();
4023
4024 // Get the initial and previous values of the scalar recurrence.
4025 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4026 auto *Previous = Phi->getIncomingValueForBlock(Latch);
4027
4028 // Create a vector from the initial value.
4029 auto *VectorInit = ScalarInit;
4030 if (VF > 1) {
4031 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
4032 VectorInit = Builder.CreateInsertElement(
4033 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4034 Builder.getInt32(VF - 1), "vector.recur.init");
4035 }
4036
4037 // We constructed a temporary phi node in the first phase of vectorization.
4038 // This phi node will eventually be deleted.
4039 VectorParts &PhiParts = VectorLoopValueMap.getVector(Phi);
4040 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
4041
4042 // Create a phi node for the new recurrence. The current value will either be
4043 // the initial value inserted into a vector or loop-varying vector value.
4044 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4045 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4046
4047 // Get the vectorized previous value.
4048 auto &PreviousParts = getVectorValue(Previous);
4049
4050 // Set the insertion point after the previous value if it is an instruction.
4051 // Note that the previous value may have been constant-folded so it is not
4052 // guaranteed to be an instruction in the vector loop.
4053 if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousParts[UF - 1]))
4054 Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt());
4055 else
4056 Builder.SetInsertPoint(
4057 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
4058
4059 // We will construct a vector for the recurrence by combining the values for
4060 // the current and previous iterations. This is the required shuffle mask.
4061 SmallVector<Constant *, 8> ShuffleMask(VF);
4062 ShuffleMask[0] = Builder.getInt32(VF - 1);
4063 for (unsigned I = 1; I < VF; ++I)
4064 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4065
4066 // The vector from which to take the initial value for the current iteration
4067 // (actual or unrolled). Initially, this is the vector phi node.
4068 Value *Incoming = VecPhi;
4069
4070 // Shuffle the current and previous vector and update the vector parts.
4071 for (unsigned Part = 0; Part < UF; ++Part) {
4072 auto *Shuffle =
4073 VF > 1
4074 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
4075 ConstantVector::get(ShuffleMask))
4076 : Incoming;
4077 PhiParts[Part]->replaceAllUsesWith(Shuffle);
4078 cast<Instruction>(PhiParts[Part])->eraseFromParent();
4079 PhiParts[Part] = Shuffle;
4080 Incoming = PreviousParts[Part];
4081 }
4082
4083 // Fix the latch value of the new recurrence in the vector loop.
4084 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4085
4086 // Extract the last vector element in the middle block. This will be the
4087 // initial value for the recurrence when jumping to the scalar loop.
4088 auto *ExtractForScalar = Incoming;
4089 if (VF > 1) {
4090 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
4091 ExtractForScalar = Builder.CreateExtractElement(
4092 ExtractForScalar, Builder.getInt32(VF - 1), "vector.recur.extract");
4093 }
4094 // Extract the second last element in the middle block if the
4095 // Phi is used outside the loop. We need to extract the phi itself
4096 // and not the last element (the phi update in the current iteration). This
4097 // will be the value when jumping to the exit block from the LoopMiddleBlock,
4098 // when the scalar loop is not run at all.
4099 Value *ExtractForPhiUsedOutsideLoop = nullptr;
4100 if (VF > 1)
4101 ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement(
4102 Incoming, Builder.getInt32(VF - 2), "vector.recur.extract.for.phi");
4103 // When loop is unrolled without vectorizing, initialize
4104 // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value of
4105 // `Incoming`. This is analogous to the vectorized case above: extracting the
4106 // second last element when VF > 1.
4107 else if (UF > 1)
4108 ExtractForPhiUsedOutsideLoop = PreviousParts[UF - 2];
4109
4110 // Fix the initial value of the original recurrence in the scalar loop.
4111 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
4112 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4113 for (auto *BB : predecessors(LoopScalarPreHeader)) {
4114 auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit;
4115 Start->addIncoming(Incoming, BB);
4116 }
4117
4118 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
4119 Phi->setName("scalar.recur");
4120
4121 // Finally, fix users of the recurrence outside the loop. The users will need
4122 // either the last value of the scalar recurrence or the last value of the
4123 // vector recurrence we extracted in the middle block. Since the loop is in
4124 // LCSSA form, we just need to find the phi node for the original scalar
4125 // recurrence in the exit block, and then add an edge for the middle block.
4126 for (auto &I : *LoopExitBlock) {
4127 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4128 if (!LCSSAPhi)
4129 break;
4130 if (LCSSAPhi->getIncomingValue(0) == Phi) {
4131 LCSSAPhi->addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock);
4132 break;
4133 }
4134 }
4135}
4136
4137void InnerLoopVectorizer::fixReduction(PHINode *Phi) {
4138 Constant *Zero = Builder.getInt32(0);
4139
4140 // Get it's reduction variable descriptor.
4141 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4142, __PRETTY_FUNCTION__))
4142 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4142, __PRETTY_FUNCTION__))
;
4143 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
4144
4145 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
4146 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
4147 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
4148 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
4149 RdxDesc.getMinMaxRecurrenceKind();
4150 setDebugLocFromInst(Builder, ReductionStartValue);
4151
4152 // We need to generate a reduction vector from the incoming scalar.
4153 // To do so, we need to generate the 'identity' vector and override
4154 // one of the elements with the incoming scalar reduction. We need
4155 // to do it in the vector-loop preheader.
4156 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
4157
4158 // This is the vector-clone of the value that leaves the loop.
4159 const VectorParts &VectorExit = getVectorValue(LoopExitInst);
4160 Type *VecTy = VectorExit[0]->getType();
4161
4162 // Find the reduction identity variable. Zero for addition, or, xor,
4163 // one for multiplication, -1 for And.
4164 Value *Identity;
4165 Value *VectorStart;
4166 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
4167 RK == RecurrenceDescriptor::RK_FloatMinMax) {
4168 // MinMax reduction have the start value as their identify.
4169 if (VF == 1) {
4170 VectorStart = Identity = ReductionStartValue;
4171 } else {
4172 VectorStart = Identity =
4173 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
4174 }
4175 } else {
4176 // Handle other reduction kinds:
4177 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
4178 RK, VecTy->getScalarType());
4179 if (VF == 1) {
4180 Identity = Iden;
4181 // This vector is the Identity vector where the first element is the
4182 // incoming scalar reduction.
4183 VectorStart = ReductionStartValue;
4184 } else {
4185 Identity = ConstantVector::getSplat(VF, Iden);
4186
4187 // This vector is the Identity vector where the first element is the
4188 // incoming scalar reduction.
4189 VectorStart =
4190 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
4191 }
4192 }
4193
4194 // Fix the vector-loop phi.
4195
4196 // Reductions do not have to start at zero. They can start with
4197 // any loop invariant values.
4198 const VectorParts &VecRdxPhi = getVectorValue(Phi);
4199 BasicBlock *Latch = OrigLoop->getLoopLatch();
4200 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
4201 const VectorParts &Val = getVectorValue(LoopVal);
4202 for (unsigned part = 0; part < UF; ++part) {
4203 // Make sure to add the reduction stat value only to the
4204 // first unroll part.
4205 Value *StartVal = (part == 0) ? VectorStart : Identity;
4206 cast<PHINode>(VecRdxPhi[part])
4207 ->addIncoming(StartVal, LoopVectorPreHeader);
4208 cast<PHINode>(VecRdxPhi[part])
4209 ->addIncoming(Val[part], LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4210 }
4211
4212 // Before each round, move the insertion point right between
4213 // the PHIs and the values we are going to write.
4214 // This allows us to write both PHINodes and the extractelement
4215 // instructions.
4216 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4217
4218 VectorParts &RdxParts = VectorLoopValueMap.getVector(LoopExitInst);
4219 setDebugLocFromInst(Builder, LoopExitInst);
4220
4221 // If the vector reduction can be performed in a smaller type, we truncate
4222 // then extend the loop exit value to enable InstCombine to evaluate the
4223 // entire expression in the smaller type.
4224 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
4225 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
4226 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
4227 for (unsigned part = 0; part < UF; ++part) {
4228 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4229 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4230 : Builder.CreateZExt(Trunc, VecTy);
4231 for (Value::user_iterator UI = RdxParts[part]->user_begin();
4232 UI != RdxParts[part]->user_end();)
4233 if (*UI != Trunc) {
4234 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
4235 RdxParts[part] = Extnd;
4236 } else {
4237 ++UI;
4238 }
4239 }
4240 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4241 for (unsigned part = 0; part < UF; ++part)
4242 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4243 }
4244
4245 // Reduce all of the unrolled parts into a single vector.
4246 Value *ReducedPartRdx = RdxParts[0];
4247 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
4248 setDebugLocFromInst(Builder, ReducedPartRdx);
4249 for (unsigned part = 1; part < UF; ++part) {
4250 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4251 // Floating point operations had to be 'fast' to enable the reduction.
4252 ReducedPartRdx = addFastMathFlag(
4253 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
4254 ReducedPartRdx, "bin.rdx"));
4255 else
4256 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4257 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
4258 }
4259
4260 if (VF > 1) {
4261 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
4262 // and vector ops, reducing the set of values being computed by half each
4263 // round.
4264 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4265, __PRETTY_FUNCTION__))
4265 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4265, __PRETTY_FUNCTION__))
;
4266 Value *TmpVec = ReducedPartRdx;
4267 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
4268 for (unsigned i = VF; i != 1; i >>= 1) {
4269 // Move the upper half of the vector to the lower half.
4270 for (unsigned j = 0; j != i / 2; ++j)
4271 ShuffleMask[j] = Builder.getInt32(i / 2 + j);
4272
4273 // Fill the rest of the mask with undef.
4274 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
4275 UndefValue::get(Builder.getInt32Ty()));
4276
4277 Value *Shuf = Builder.CreateShuffleVector(
4278 TmpVec, UndefValue::get(TmpVec->getType()),
4279 ConstantVector::get(ShuffleMask), "rdx.shuf");
4280
4281 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4282 // Floating point operations had to be 'fast' to enable the reduction.
4283 TmpVec = addFastMathFlag(Builder.CreateBinOp(
4284 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
4285 else
4286 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
4287 TmpVec, Shuf);
4288 }
4289
4290 // The result is in the first element of the vector.
4291 ReducedPartRdx =
4292 Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4293
4294 // If the reduction can be performed in a smaller type, we need to extend
4295 // the reduction to the wider type before we branch to the original loop.
4296 if (Phi->getType() != RdxDesc.getRecurrenceType())
4297 ReducedPartRdx =
4298 RdxDesc.isSigned()
4299 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4300 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4301 }
4302
4303 // Create a phi node that merges control-flow from the backedge-taken check
4304 // block and the middle block.
4305 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4306 LoopScalarPreHeader->getTerminator());
4307 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4308 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4309 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4310
4311 // Now, we need to fix the users of the reduction variable
4312 // inside and outside of the scalar remainder loop.
4313 // We know that the loop is in LCSSA form. We need to update the
4314 // PHI nodes in the exit blocks.
4315 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4316 LEE = LoopExitBlock->end();
4317 LEI != LEE; ++LEI) {
4318 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4319 if (!LCSSAPhi)
4320 break;
4321
4322 // All PHINodes need to have a single entry edge, or two if
4323 // we already fixed them.
4324 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4324, __PRETTY_FUNCTION__))
;
4325
4326 // We found a reduction value exit-PHI. Update it with the
4327 // incoming bypass edge.
4328 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst)
4329 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4330 } // end of the LCSSA phi scan.
4331
4332 // Fix the scalar loop reduction variable with the incoming reduction sum
4333 // from the vector body and from the backedge value.
4334 int IncomingEdgeBlockIdx =
4335 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
4336 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4336, __PRETTY_FUNCTION__))
;
4337 // Pick the other block.
4338 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4339 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4340 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4341}
4342
4343void InnerLoopVectorizer::fixLCSSAPHIs() {
4344 for (Instruction &LEI : *LoopExitBlock) {
4345 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4346 if (!LCSSAPhi)
4347 break;
4348 if (LCSSAPhi->getNumIncomingValues() == 1)
4349 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
4350 LoopMiddleBlock);
4351 }
4352}
4353
4354void InnerLoopVectorizer::collectTriviallyDeadInstructions(
4355 SmallPtrSetImpl<Instruction *> &DeadInstructions) {
4356 BasicBlock *Latch = OrigLoop->getLoopLatch();
4357
4358 // We create new control-flow for the vectorized loop, so the original
4359 // condition will be dead after vectorization if it's only used by the
4360 // branch.
4361 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
4362 if (Cmp && Cmp->hasOneUse())
4363 DeadInstructions.insert(Cmp);
4364
4365 // We create new "steps" for induction variable updates to which the original
4366 // induction variables map. An original update instruction will be dead if
4367 // all its users except the induction variable are dead.
4368 for (auto &Induction : *Legal->getInductionVars()) {
4369 PHINode *Ind = Induction.first;
4370 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
4371 if (all_of(IndUpdate->users(), [&](User *U) -> bool {
4372 return U == Ind || DeadInstructions.count(cast<Instruction>(U));
4373 }))
4374 DeadInstructions.insert(IndUpdate);
4375 }
4376}
4377
4378void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) {
4379
4380 // The basic block and loop containing the predicated instruction.
4381 auto *PredBB = PredInst->getParent();
4382 auto *VectorLoop = LI->getLoopFor(PredBB);
4383
4384 // Initialize a worklist with the operands of the predicated instruction.
4385 SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4386
4387 // Holds instructions that we need to analyze again. An instruction may be
4388 // reanalyzed if we don't yet know if we can sink it or not.
4389 SmallVector<Instruction *, 8> InstsToReanalyze;
4390
4391 // Returns true if a given use occurs in the predicated block. Phi nodes use
4392 // their operands in their corresponding predecessor blocks.
4393 auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4394 auto *I = cast<Instruction>(U.getUser());
4395 BasicBlock *BB = I->getParent();
4396 if (auto *Phi = dyn_cast<PHINode>(I))
4397 BB = Phi->getIncomingBlock(
4398 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4399 return BB == PredBB;
4400 };
4401
4402 // Iteratively sink the scalarized operands of the predicated instruction
4403 // into the block we created for it. When an instruction is sunk, it's
4404 // operands are then added to the worklist. The algorithm ends after one pass
4405 // through the worklist doesn't sink a single instruction.
4406 bool Changed;
4407 do {
4408
4409 // Add the instructions that need to be reanalyzed to the worklist, and
4410 // reset the changed indicator.
4411 Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4412 InstsToReanalyze.clear();
4413 Changed = false;
4414
4415 while (!Worklist.empty()) {
4416 auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4417
4418 // We can't sink an instruction if it is a phi node, is already in the
4419 // predicated block, is not in the loop, or may have side effects.
4420 if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4421 !VectorLoop->contains(I) || I->mayHaveSideEffects())
4422 continue;
4423
4424 // It's legal to sink the instruction if all its uses occur in the
4425 // predicated block. Otherwise, there's nothing to do yet, and we may
4426 // need to reanalyze the instruction.
4427 if (!all_of(I->uses(), isBlockOfUsePredicated)) {
4428 InstsToReanalyze.push_back(I);
4429 continue;
4430 }
4431
4432 // Move the instruction to the beginning of the predicated block, and add
4433 // it's operands to the worklist.
4434 I->moveBefore(&*PredBB->getFirstInsertionPt());
4435 Worklist.insert(I->op_begin(), I->op_end());
4436
4437 // The sinking may have enabled other instructions to be sunk, so we will
4438 // need to iterate.
4439 Changed = true;
4440 }
4441 } while (Changed);
4442}
4443
4444void InnerLoopVectorizer::predicateInstructions() {
4445
4446 // For each instruction I marked for predication on value C, split I into its
4447 // own basic block to form an if-then construct over C. Since I may be fed by
4448 // an extractelement instruction or other scalar operand, we try to
4449 // iteratively sink its scalar operands into the predicated block. If I feeds
4450 // an insertelement instruction, we try to move this instruction into the
4451 // predicated block as well. For non-void types, a phi node will be created
4452 // for the resulting value (either vector or scalar).
4453 //
4454 // So for some predicated instruction, e.g. the conditional sdiv in:
4455 //
4456 // for.body:
4457 // ...
4458 // %add = add nsw i32 %mul, %0
4459 // %cmp5 = icmp sgt i32 %2, 7
4460 // br i1 %cmp5, label %if.then, label %if.end
4461 //
4462 // if.then:
4463 // %div = sdiv i32 %0, %1
4464 // br label %if.end
4465 //
4466 // if.end:
4467 // %x.0 = phi i32 [ %div, %if.then ], [ %add, %for.body ]
4468 //
4469 // the sdiv at this point is scalarized and if-converted using a select.
4470 // The inactive elements in the vector are not used, but the predicated
4471 // instruction is still executed for all vector elements, essentially:
4472 //
4473 // vector.body:
4474 // ...
4475 // %17 = add nsw <2 x i32> %16, %wide.load
4476 // %29 = extractelement <2 x i32> %wide.load, i32 0
4477 // %30 = extractelement <2 x i32> %wide.load51, i32 0
4478 // %31 = sdiv i32 %29, %30
4479 // %32 = insertelement <2 x i32> undef, i32 %31, i32 0
4480 // %35 = extractelement <2 x i32> %wide.load, i32 1
4481 // %36 = extractelement <2 x i32> %wide.load51, i32 1
4482 // %37 = sdiv i32 %35, %36
4483 // %38 = insertelement <2 x i32> %32, i32 %37, i32 1
4484 // %predphi = select <2 x i1> %26, <2 x i32> %38, <2 x i32> %17
4485 //
4486 // Predication will now re-introduce the original control flow to avoid false
4487 // side-effects by the sdiv instructions on the inactive elements, yielding
4488 // (after cleanup):
4489 //
4490 // vector.body:
4491 // ...
4492 // %5 = add nsw <2 x i32> %4, %wide.load
4493 // %8 = icmp sgt <2 x i32> %wide.load52, <i32 7, i32 7>
4494 // %9 = extractelement <2 x i1> %8, i32 0
4495 // br i1 %9, label %pred.sdiv.if, label %pred.sdiv.continue
4496 //
4497 // pred.sdiv.if:
4498 // %10 = extractelement <2 x i32> %wide.load, i32 0
4499 // %11 = extractelement <2 x i32> %wide.load51, i32 0
4500 // %12 = sdiv i32 %10, %11
4501 // %13 = insertelement <2 x i32> undef, i32 %12, i32 0
4502 // br label %pred.sdiv.continue
4503 //
4504 // pred.sdiv.continue:
4505 // %14 = phi <2 x i32> [ undef, %vector.body ], [ %13, %pred.sdiv.if ]
4506 // %15 = extractelement <2 x i1> %8, i32 1
4507 // br i1 %15, label %pred.sdiv.if54, label %pred.sdiv.continue55
4508 //
4509 // pred.sdiv.if54:
4510 // %16 = extractelement <2 x i32> %wide.load, i32 1
4511 // %17 = extractelement <2 x i32> %wide.load51, i32 1
4512 // %18 = sdiv i32 %16, %17
4513 // %19 = insertelement <2 x i32> %14, i32 %18, i32 1
4514 // br label %pred.sdiv.continue55
4515 //
4516 // pred.sdiv.continue55:
4517 // %20 = phi <2 x i32> [ %14, %pred.sdiv.continue ], [ %19, %pred.sdiv.if54 ]
4518 // %predphi = select <2 x i1> %8, <2 x i32> %20, <2 x i32> %5
4519
4520 for (auto KV : PredicatedInstructions) {
4521 BasicBlock::iterator I(KV.first);
4522 BasicBlock *Head = I->getParent();
4523 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
4524 /*BranchWeights=*/nullptr, DT, LI);
4525 I->moveBefore(T);
4526 sinkScalarOperands(&*I);
4527
4528 BasicBlock *PredicatedBlock = I->getParent();
4529 Twine BBNamePrefix = Twine("pred.") + I->getOpcodeName();
4530 PredicatedBlock->setName(BBNamePrefix + ".if");
4531 PredicatedBlock->getSingleSuccessor()->setName(BBNamePrefix + ".continue");
4532
4533 // If the instruction is non-void create a Phi node at reconvergence point.
4534 if (!I->getType()->isVoidTy()) {
4535 Value *IncomingTrue = nullptr;
4536 Value *IncomingFalse = nullptr;
4537
4538 if (I->hasOneUse() && isa<InsertElementInst>(*I->user_begin())) {
4539 // If the predicated instruction is feeding an insert-element, move it
4540 // into the Then block; Phi node will be created for the vector.
4541 InsertElementInst *IEI = cast<InsertElementInst>(*I->user_begin());
4542 IEI->moveBefore(T);
4543 IncomingTrue = IEI; // the new vector with the inserted element.
4544 IncomingFalse = IEI->getOperand(0); // the unmodified vector
4545 } else {
4546 // Phi node will be created for the scalar predicated instruction.
4547 IncomingTrue = &*I;
4548 IncomingFalse = UndefValue::get(I->getType());
4549 }
4550
4551 BasicBlock *PostDom = I->getParent()->getSingleSuccessor();
4552 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4552, __PRETTY_FUNCTION__))
;
4553 PHINode *Phi =
4554 PHINode::Create(IncomingTrue->getType(), 2, "", &PostDom->front());
4555 IncomingTrue->replaceAllUsesWith(Phi);
4556 Phi->addIncoming(IncomingFalse, Head);
4557 Phi->addIncoming(IncomingTrue, I->getParent());
4558 }
4559 }
4560
4561 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (false
)
;
4562}
4563
4564InnerLoopVectorizer::VectorParts
4565InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
4566 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4566, __PRETTY_FUNCTION__))
;
4567
4568 // Look for cached value.
4569 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
4570 EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge);
4571 if (ECEntryIt != EdgeMaskCache.end())
4572 return ECEntryIt->second;
4573
4574 VectorParts SrcMask = createBlockInMask(Src);
4575
4576 // The terminator has to be a branch inst!
4577 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
4578 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4578, __PRETTY_FUNCTION__))
;
4579
4580 if (BI->isConditional()) {
4581 VectorParts EdgeMask = getVectorValue(BI->getCondition());
4582
4583 if (BI->getSuccessor(0) != Dst)
4584 for (unsigned part = 0; part < UF; ++part)
4585 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4586
4587 for (unsigned part = 0; part < UF; ++part)
4588 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4589
4590 EdgeMaskCache[Edge] = EdgeMask;
4591 return EdgeMask;
4592 }
4593
4594 EdgeMaskCache[Edge] = SrcMask;
4595 return SrcMask;
4596}
4597
4598InnerLoopVectorizer::VectorParts
4599InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4600 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4600, __PRETTY_FUNCTION__))
;
4601
4602 // Look for cached value.
4603 BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB);
4604 if (BCEntryIt != BlockMaskCache.end())
4605 return BCEntryIt->second;
4606
4607 // Loop incoming mask is all-one.
4608 if (OrigLoop->getHeader() == BB) {
4609 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4610 const VectorParts &BlockMask = getVectorValue(C);
4611 BlockMaskCache[BB] = BlockMask;
4612 return BlockMask;
4613 }
4614
4615 // This is the block mask. We OR all incoming edges, and with zero.
4616 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4617 VectorParts BlockMask = getVectorValue(Zero);
4618
4619 // For each pred:
4620 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4621 VectorParts EM = createEdgeMask(*it, BB);
4622 for (unsigned part = 0; part < UF; ++part)
4623 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4624 }
4625
4626 BlockMaskCache[BB] = BlockMask;
4627 return BlockMask;
4628}
4629
4630void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
4631 unsigned VF) {
4632 PHINode *P = cast<PHINode>(PN);
4633 // In order to support recurrences we need to be able to vectorize Phi nodes.
4634 // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4635 // stage #1: We create a new vector PHI node with no incoming edges. We'll use
4636 // this value when we vectorize all of the instructions that use the PHI.
4637 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4638 VectorParts Entry(UF);
4639 for (unsigned part = 0; part < UF; ++part) {
4640 // This is phase one of vectorizing PHIs.
4641 Type *VecTy =
4642 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4643 Entry[part] = PHINode::Create(
4644 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4645 }
4646 VectorLoopValueMap.initVector(P, Entry);
4647 return;
4648 }
4649
4650 setDebugLocFromInst(Builder, P);
4651 // Check for PHI nodes that are lowered to vector selects.
4652 if (P->getParent() != OrigLoop->getHeader()) {
4653 // We know that all PHIs in non-header blocks are converted into
4654 // selects, so we don't have to worry about the insertion order and we
4655 // can just use the builder.
4656 // At this point we generate the predication tree. There may be
4657 // duplications since this is a simple recursive scan, but future
4658 // optimizations will clean it up.
4659
4660 unsigned NumIncoming = P->getNumIncomingValues();
4661
4662 // Generate a sequence of selects of the form:
4663 // SELECT(Mask3, In3,
4664 // SELECT(Mask2, In2,
4665 // ( ...)))
4666 VectorParts Entry(UF);
4667 for (unsigned In = 0; In < NumIncoming; In++) {
4668 VectorParts Cond =
4669 createEdgeMask(P->getIncomingBlock(In), P->getParent());
4670 const VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
4671
4672 for (unsigned part = 0; part < UF; ++part) {
4673 // We might have single edge PHIs (blocks) - use an identity
4674 // 'select' for the first PHI operand.
4675 if (In == 0)
4676 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In0[part]);
4677 else
4678 // Select between the current value and the previous incoming edge
4679 // based on the incoming mask.
4680 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], Entry[part],
4681 "predphi");
4682 }
4683 }
4684 VectorLoopValueMap.initVector(P, Entry);
4685 return;
4686 }
4687
4688 // This PHINode must be an induction variable.
4689 // Make sure that we know about it.
4690 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4690, __PRETTY_FUNCTION__))
;
4691
4692 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4693 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4694
4695 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4696 // which can be found from the original scalar operations.
4697 switch (II.getKind()) {
4698 case InductionDescriptor::IK_NoInduction:
4699 llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4699)
;
4700 case InductionDescriptor::IK_IntInduction:
4701 case InductionDescriptor::IK_FpInduction:
4702 return widenIntOrFpInduction(P);
4703 case InductionDescriptor::IK_PtrInduction: {
4704 // Handle the pointer induction variable case.
4705 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4705, __PRETTY_FUNCTION__))
;
4706 // This is the normalized GEP that starts counting at zero.
4707 Value *PtrInd = Induction;
4708 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4709 // Determine the number of scalars we need to generate for each unroll
4710 // iteration. If the instruction is uniform, we only need to generate the
4711 // first lane. Otherwise, we generate all VF values.
4712 unsigned Lanes = Cost->isUniformAfterVectorization(P, VF) ? 1 : VF;
4713 // These are the scalar results. Notice that we don't generate vector GEPs
4714 // because scalar GEPs result in better code.
4715 ScalarParts Entry(UF);
4716 for (unsigned Part = 0; Part < UF; ++Part) {
4717 Entry[Part].resize(VF);
4718 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4719 Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4720 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4721 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4722 SclrGep->setName("next.gep");
4723 Entry[Part][Lane] = SclrGep;
4724 }
4725 }
4726 VectorLoopValueMap.initScalar(P, Entry);
4727 return;
4728 }
4729 }
4730}
4731
4732/// A helper function for checking whether an integer division-related
4733/// instruction may divide by zero (in which case it must be predicated if
4734/// executed conditionally in the scalar code).
4735/// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4736/// Non-zero divisors that are non compile-time constants will not be
4737/// converted into multiplication, so we will still end up scalarizing
4738/// the division, but can do so w/o predication.
4739static bool mayDivideByZero(Instruction &I) {
4740 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4744, __PRETTY_FUNCTION__))
4741 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4744, __PRETTY_FUNCTION__))
4742 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4744, __PRETTY_FUNCTION__))
4743 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4744, __PRETTY_FUNCTION__))
4744 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4744, __PRETTY_FUNCTION__))
;
4745 Value *Divisor = I.getOperand(1);
4746 auto *CInt = dyn_cast<ConstantInt>(Divisor);
4747 return !CInt || CInt->isZero();
4748}
4749
4750void InnerLoopVectorizer::vectorizeInstruction(Instruction &I) {
4751 // Scalarize instructions that should remain scalar after vectorization.
4752 if (VF > 1 &&
4753 !(isa<BranchInst>(&I) || isa<PHINode>(&I) || isa<DbgInfoIntrinsic>(&I)) &&
4754 shouldScalarizeInstruction(&I)) {
4755 scalarizeInstruction(&I, Legal->isScalarWithPredication(&I));
4756 return;
4757 }
4758
4759 switch (I.getOpcode()) {
4760 case Instruction::Br:
4761 // Nothing to do for PHIs and BR, since we already took care of the
4762 // loop control flow instructions.
4763 break;
4764 case Instruction::PHI: {
4765 // Vectorize PHINodes.
4766 widenPHIInstruction(&I, UF, VF);
4767 break;
4768 } // End of PHI.
4769 case Instruction::GetElementPtr: {
4770 // Construct a vector GEP by widening the operands of the scalar GEP as
4771 // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP
4772 // results in a vector of pointers when at least one operand of the GEP
4773 // is vector-typed. Thus, to keep the representation compact, we only use
4774 // vector-typed operands for loop-varying values.
4775 auto *GEP = cast<GetElementPtrInst>(&I);
4776 VectorParts Entry(UF);
4777
4778 if (VF > 1 && OrigLoop->hasLoopInvariantOperands(GEP)) {
4779 // If we are vectorizing, but the GEP has only loop-invariant operands,
4780 // the GEP we build (by only using vector-typed operands for
4781 // loop-varying values) would be a scalar pointer. Thus, to ensure we
4782 // produce a vector of pointers, we need to either arbitrarily pick an
4783 // operand to broadcast, or broadcast a clone of the original GEP.
4784 // Here, we broadcast a clone of the original.
4785 //
4786 // TODO: If at some point we decide to scalarize instructions having
4787 // loop-invariant operands, this special case will no longer be
4788 // required. We would add the scalarization decision to
4789 // collectLoopScalars() and teach getVectorValue() to broadcast
4790 // the lane-zero scalar value.
4791 auto *Clone = Builder.Insert(GEP->clone());
4792 for (unsigned Part = 0; Part < UF; ++Part)
4793 Entry[Part] = Builder.CreateVectorSplat(VF, Clone);
4794 } else {
4795 // If the GEP has at least one loop-varying operand, we are sure to
4796 // produce a vector of pointers. But if we are only unrolling, we want
4797 // to produce a scalar GEP for each unroll part. Thus, the GEP we
4798 // produce with the code below will be scalar (if VF == 1) or vector
4799 // (otherwise). Note that for the unroll-only case, we still maintain
4800 // values in the vector mapping with initVector, as we do for other
4801 // instructions.
4802 for (unsigned Part = 0; Part < UF; ++Part) {
4803
4804 // The pointer operand of the new GEP. If it's loop-invariant, we
4805 // won't broadcast it.
4806 auto *Ptr = OrigLoop->isLoopInvariant(GEP->getPointerOperand())
4807 ? GEP->getPointerOperand()
4808 : getVectorValue(GEP->getPointerOperand())[Part];
4809
4810 // Collect all the indices for the new GEP. If any index is
4811 // loop-invariant, we won't broadcast it.
4812 SmallVector<Value *, 4> Indices;
4813 for (auto &U : make_range(GEP->idx_begin(), GEP->idx_end())) {
4814 if (OrigLoop->isLoopInvariant(U.get()))
4815 Indices.push_back(U.get());
4816 else
4817 Indices.push_back(getVectorValue(U.get())[Part]);
4818 }
4819
4820 // Create the new GEP. Note that this GEP may be a scalar if VF == 1,
4821 // but it should be a vector, otherwise.
4822 auto *NewGEP = GEP->isInBounds()
4823 ? Builder.CreateInBoundsGEP(Ptr, Indices)
4824 : Builder.CreateGEP(Ptr, Indices);
4825 assert((VF == 1 || NewGEP->getType()->isVectorTy()) &&(((VF == 1 || NewGEP->getType()->isVectorTy()) &&
"NewGEP is not a pointer vector") ? static_cast<void> (
0) : __assert_fail ("(VF == 1 || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4826, __PRETTY_FUNCTION__))
4826 "NewGEP is not a pointer vector")(((VF == 1 || NewGEP->getType()->isVectorTy()) &&
"NewGEP is not a pointer vector") ? static_cast<void> (
0) : __assert_fail ("(VF == 1 || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 4826, __PRETTY_FUNCTION__))
;
4827 Entry[Part] = NewGEP;
4828 }
4829 }
4830
4831 VectorLoopValueMap.initVector(&I, Entry);
4832 addMetadata(Entry, GEP);
4833 break;
4834 }
4835 case Instruction::UDiv:
4836 case Instruction::SDiv:
4837 case Instruction::SRem:
4838 case Instruction::URem:
4839 // Scalarize with predication if this instruction may divide by zero and
4840 // block execution is conditional, otherwise fallthrough.
4841 if (Legal->isScalarWithPredication(&I)) {
4842 scalarizeInstruction(&I, true);
4843 break;
4844 }
4845 case Instruction::Add:
4846 case Instruction::FAdd:
4847 case Instruction::Sub:
4848 case Instruction::FSub:
4849 case Instruction::Mul:
4850 case Instruction::FMul:
4851 case Instruction::FDiv:
4852 case Instruction::FRem:
4853 case Instruction::Shl:
4854 case Instruction::LShr:
4855 case Instruction::AShr:
4856 case Instruction::And:
4857 case Instruction::Or:
4858 case Instruction::Xor: {
4859 // Just widen binops.
4860 auto *BinOp = cast<BinaryOperator>(&I);
4861 setDebugLocFromInst(Builder, BinOp);
4862 const VectorParts &A = getVectorValue(BinOp->getOperand(0));
4863 const VectorParts &B = getVectorValue(BinOp->getOperand(1));
4864
4865 // Use this vector value for all users of the original instruction.
4866 VectorParts Entry(UF);
4867 for (unsigned Part = 0; Part < UF; ++Part) {
4868 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4869
4870 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4871 VecOp->copyIRFlags(BinOp);
4872
4873 Entry[Part] = V;
4874 }
4875
4876 VectorLoopValueMap.initVector(&I, Entry);
4877 addMetadata(Entry, BinOp);
4878 break;
4879 }
4880 case Instruction::Select: {
4881 // Widen selects.
4882 // If the selector is loop invariant we can create a select
4883 // instruction with a scalar condition. Otherwise, use vector-select.
4884 auto *SE = PSE.getSE();
4885 bool InvariantCond =
4886 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4887 setDebugLocFromInst(Builder, &I);
4888
4889 // The condition can be loop invariant but still defined inside the
4890 // loop. This means that we can't just use the original 'cond' value.
4891 // We have to take the 'vectorized' value and pick the first lane.
4892 // Instcombine will make this a no-op.
4893 const VectorParts &Cond = getVectorValue(I.getOperand(0));
4894 const VectorParts &Op0 = getVectorValue(I.getOperand(1));
4895 const VectorParts &Op1 = getVectorValue(I.getOperand(2));
4896
4897 auto *ScalarCond = getScalarValue(I.getOperand(0), 0, 0);
4898
4899 VectorParts Entry(UF);
4900 for (unsigned Part = 0; Part < UF; ++Part) {
4901 Entry[Part] = Builder.CreateSelect(
4902 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4903 }
4904
4905 VectorLoopValueMap.initVector(&I, Entry);
4906 addMetadata(Entry, &I);
4907 break;
4908 }
4909
4910 case Instruction::ICmp:
4911 case Instruction::FCmp: {
4912 // Widen compares. Generate vector compares.
4913 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4914 auto *Cmp = dyn_cast<CmpInst>(&I);
4915 setDebugLocFromInst(Builder, Cmp);
4916 const VectorParts &A = getVectorValue(Cmp->getOperand(0));
4917 const VectorParts &B = getVectorValue(Cmp->getOperand(1));
4918 VectorParts Entry(UF);
4919 for (unsigned Part = 0; Part < UF; ++Part) {
4920 Value *C = nullptr;
4921 if (FCmp) {
4922 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4923 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4924 } else {
4925 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4926 }
4927 Entry[Part] = C;
4928 }
4929
4930 VectorLoopValueMap.initVector(&I, Entry);
4931 addMetadata(Entry, &I);
4932 break;
4933 }
4934
4935 case Instruction::Store:
4936 case Instruction::Load:
4937 vectorizeMemoryInstruction(&I);
4938 break;
4939 case Instruction::ZExt:
4940 case Instruction::SExt:
4941 case Instruction::FPToUI:
4942 case Instruction::FPToSI:
4943 case Instruction::FPExt:
4944 case Instruction::PtrToInt:
4945 case Instruction::IntToPtr:
4946 case Instruction::SIToFP:
4947 case Instruction::UIToFP:
4948 case Instruction::Trunc:
4949 case Instruction::FPTrunc:
4950 case Instruction::BitCast: {
4951 auto *CI = dyn_cast<CastInst>(&I);
4952 setDebugLocFromInst(Builder, CI);
4953
4954 // Optimize the special case where the source is a constant integer
4955 // induction variable. Notice that we can only optimize the 'trunc' case
4956 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4957 // (c) other casts depend on pointer size.
4958 if (Cost->isOptimizableIVTruncate(CI, VF)) {
4959 widenIntOrFpInduction(cast<PHINode>(CI->getOperand(0)),
4960 cast<TruncInst>(CI));
4961 break;
4962 }
4963
4964 /// Vectorize casts.
4965 Type *DestTy =
4966 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4967
4968 const VectorParts &A = getVectorValue(CI->getOperand(0));
4969 VectorParts Entry(UF);
4970 for (unsigned Part = 0; Part < UF; ++Part)
4971 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4972 VectorLoopValueMap.initVector(&I, Entry);
4973 addMetadata(Entry, &I);
4974 break;
4975 }
4976
4977 case Instruction::Call: {
4978 // Ignore dbg intrinsics.
4979 if (isa<DbgInfoIntrinsic>(I))
4980 break;
4981 setDebugLocFromInst(Builder, &I);
4982
4983 Module *M = I.getParent()->getParent()->getParent();
4984 auto *CI = cast<CallInst>(&I);
4985
4986 StringRef FnName = CI->getCalledFunction()->getName();
4987 Function *F = CI->getCalledFunction();
4988 Type *RetTy = ToVectorTy(CI->getType(), VF);
4989 SmallVector<Type *, 4> Tys;
4990 for (Value *ArgOperand : CI->arg_operands())
4991 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4992
4993 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4994 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4995 ID == Intrinsic::lifetime_start)) {
4996 scalarizeInstruction(&I);
4997 break;
4998 }
4999 // The flag shows whether we use Intrinsic or a usual Call for vectorized
5000 // version of the instruction.
5001 // Is it beneficial to perform intrinsic call compared to lib call?
5002 bool NeedToScalarize;
5003 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
5004 bool UseVectorIntrinsic =
5005 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
5006 if (!UseVectorIntrinsic && NeedToScalarize) {
5007 scalarizeInstruction(&I);
5008 break;
5009 }
5010
5011 VectorParts Entry(UF);
5012 for (unsigned Part = 0; Part < UF; ++Part) {
5013 SmallVector<Value *, 4> Args;
5014 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
5015 Value *Arg = CI->getArgOperand(i);
5016 // Some intrinsics have a scalar argument - don't replace it with a
5017 // vector.
5018 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
5019 const VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
5020 Arg = VectorArg[Part];
5021 }
5022 Args.push_back(Arg);
5023 }
5024
5025 Function *VectorF;
5026 if (UseVectorIntrinsic) {
5027 // Use vector version of the intrinsic.
5028 Type *TysForDecl[] = {CI->getType()};
5029 if (VF > 1)
5030 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
5031 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
5032 } else {
5033 // Use vector version of the library call.
5034 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
5035 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5035, __PRETTY_FUNCTION__))
;
5036 VectorF = M->getFunction(VFnName);
5037 if (!VectorF) {
5038 // Generate a declaration
5039 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
5040 VectorF =
5041 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
5042 VectorF->copyAttributesFrom(F);
5043 }
5044 }
5045 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5045, __PRETTY_FUNCTION__))
;
5046
5047 SmallVector<OperandBundleDef, 1> OpBundles;
5048 CI->getOperandBundlesAsDefs(OpBundles);
5049 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
5050
5051 if (isa<FPMathOperator>(V))
5052 V->copyFastMathFlags(CI);
5053
5054 Entry[Part] = V;
5055 }
5056
5057 VectorLoopValueMap.initVector(&I, Entry);
5058 addMetadata(Entry, &I);
5059 break;
5060 }
5061
5062 default:
5063 // All other instructions are unsupported. Scalarize them.
5064 scalarizeInstruction(&I);
5065 break;
5066 } // end of switch.
5067}
5068
5069void InnerLoopVectorizer::updateAnalysis() {
5070 // Forget the original basic block.
5071 PSE.getSE()->forgetLoop(OrigLoop);
5072
5073 // Update the dominator tree information.
5074 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5075, __PRETTY_FUNCTION__))
5075 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5075, __PRETTY_FUNCTION__))
;
5076
5077 DT->addNewBlock(LI->getLoopFor(LoopVectorBody)->getHeader(),
5078 LoopVectorPreHeader);
5079 DT->addNewBlock(LoopMiddleBlock,
5080 LI->getLoopFor(LoopVectorBody)->getLoopLatch());
5081 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
5082 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
5083 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
5084
5085 DEBUG(DT->verifyDomTree())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { DT->verifyDomTree(); } } while (false
)
;
5086}
5087
5088/// \brief Check whether it is safe to if-convert this phi node.
5089///
5090/// Phi nodes with constant expressions that can trap are not safe to if
5091/// convert.
5092static bool canIfConvertPHINodes(BasicBlock *BB) {
5093 for (Instruction &I : *BB) {
5094 auto *Phi = dyn_cast<PHINode>(&I);
5095 if (!Phi)
5096 return true;
5097 for (Value *V : Phi->incoming_values())
5098 if (auto *C = dyn_cast<Constant>(V))
5099 if (C->canTrap())
5100 return false;
5101 }
5102 return true;
5103}
5104
5105bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
5106 if (!EnableIfConversion) {
5107 ORE->emit(createMissedAnalysis("IfConversionDisabled")
5108 << "if-conversion is disabled");
5109 return false;
5110 }
5111
5112 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5112, __PRETTY_FUNCTION__))
;
5113
5114 // A list of pointers that we can safely read and write to.
5115 SmallPtrSet<Value *, 8> SafePointes;
5116
5117 // Collect safe addresses.
5118 for (BasicBlock *BB : TheLoop->blocks()) {
5119 if (blockNeedsPredication(BB))
5120 continue;
5121
5122 for (Instruction &I : *BB)
5123 if (auto *Ptr = getPointerOperand(&I))
5124 SafePointes.insert(Ptr);
5125 }
5126
5127 // Collect the blocks that need predication.
5128 BasicBlock *Header = TheLoop->getHeader();
5129 for (BasicBlock *BB : TheLoop->blocks()) {
5130 // We don't support switch statements inside loops.
5131 if (!isa<BranchInst>(BB->getTerminator())) {
5132 ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5133 << "loop contains a switch statement");
5134 return false;
5135 }
5136
5137 // We must be able to predicate all blocks that need to be predicated.
5138 if (blockNeedsPredication(BB)) {
5139 if (!blockCanBePredicated(BB, SafePointes)) {
5140 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5141 << "control flow cannot be substituted for a select");
5142 return false;
5143 }
5144 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5145 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5146 << "control flow cannot be substituted for a select");
5147 return false;
5148 }
5149 }
5150
5151 // We can if-convert this loop.
5152 return true;
5153}
5154
5155bool LoopVectorizationLegality::canVectorize() {
5156 // We must have a loop in canonical form. Loops with indirectbr in them cannot
5157 // be canonicalized.
5158 if (!TheLoop->getLoopPreheader()) {
5159 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5160 << "loop control flow is not understood by vectorizer");
5161 return false;
5162 }
5163
5164 // FIXME: The code is currently dead, since the loop gets sent to
5165 // LoopVectorizationLegality is already an innermost loop.
5166 //
5167 // We can only vectorize innermost loops.
5168 if (!TheLoop->empty()) {
5169 ORE->emit(createMissedAnalysis("NotInnermostLoop")
5170 << "loop is not the innermost loop");
5171 return false;
5172 }
5173
5174 // We must have a single backedge.
5175 if (TheLoop->getNumBackEdges() != 1) {
5176 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5177 << "loop control flow is not understood by vectorizer");
5178 return false;
5179 }
5180
5181 // We must have a single exiting block.
5182 if (!TheLoop->getExitingBlock()) {
5183 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5184 << "loop control flow is not understood by vectorizer");
5185 return false;
5186 }
5187
5188 // We only handle bottom-tested loops, i.e. loop in which the condition is
5189 // checked at the end of each iteration. With that we can assume that all
5190 // instructions in the loop are executed the same number of times.
5191 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5192 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5193 << "loop control flow is not understood by vectorizer");
5194 return false;
5195 }
5196
5197 // We need to have a loop header.
5198 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)
5199 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found a loop: " <<
TheLoop->getHeader()->getName() << '\n'; } } while
(false)
;
5200
5201 // Check if we can if-convert non-single-bb loops.
5202 unsigned NumBlocks = TheLoop->getNumBlocks();
5203 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5204 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)
;
5205 return false;
5206 }
5207
5208 // ScalarEvolution needs to be able to find the exit count.
5209 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
5210 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
5211 ORE->emit(createMissedAnalysis("CantComputeNumberOfIterations")
5212 << "could not determine number of loop iterations");
5213 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)
;
5214 return false;
5215 }
5216
5217 // Check if we can vectorize the instructions and CFG in this loop.
5218 if (!canVectorizeInstrs()) {
5219 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)
;
5220 return false;
5221 }
5222
5223 // Go over each instruction and look at memory deps.
5224 if (!canVectorizeMemory()) {
5225 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)
;
5226 return false;
5227 }
5228
5229 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)
5230 << (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)
5231 ? " (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)
5232 : "")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)
5233 << "!\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)
;
5234
5235 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5236
5237 // If an override option has been passed in for interleaved accesses, use it.
5238 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5239 UseInterleaved = EnableInterleavedMemAccesses;
5240
5241 // Analyze interleaved memory accesses.
5242 if (UseInterleaved)
5243 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5244
5245 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5246 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5247 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5248
5249 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5250 ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5251 << "Too many SCEV assumptions need to be made and checked "
5252 << "at runtime");
5253 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)
;
5254 return false;
5255 }
5256
5257 // Okay! We can vectorize. At this point we don't have any other mem analysis
5258 // which may limit our maximum vectorization factor, so just return true with
5259 // no restrictions.
5260 return true;
5261}
5262
5263static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
5264 if (Ty->isPointerTy())
5265 return DL.getIntPtrType(Ty);
5266
5267 // It is possible that char's or short's overflow when we ask for the loop's
5268 // trip count, work around this by changing the type size.
5269 if (Ty->getScalarSizeInBits() < 32)
5270 return Type::getInt32Ty(Ty->getContext());
5271
5272 return Ty;
5273}
5274
5275static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
5276 Ty0 = convertPointerToIntegerType(DL, Ty0);
5277 Ty1 = convertPointerToIntegerType(DL, Ty1);
5278 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
5279 return Ty0;
5280 return Ty1;
5281}
5282
5283/// \brief Check that the instruction has outside loop users and is not an
5284/// identified reduction variable.
5285static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
5286 SmallPtrSetImpl<Value *> &AllowedExit) {
5287 // Reduction and Induction instructions are allowed to have exit users. All
5288 // other instructions must not have external users.
5289 if (!AllowedExit.count(Inst))
5290 // Check that all of the users of the loop are inside the BB.
5291 for (User *U : Inst->users()) {
5292 Instruction *UI = cast<Instruction>(U);
5293 // This user may be a reduction exit value.
5294 if (!TheLoop->contains(UI)) {
5295 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)
;
5296 return true;
5297 }
5298 }
5299 return false;
5300}
5301
5302void LoopVectorizationLegality::addInductionPhi(
5303 PHINode *Phi, const InductionDescriptor &ID,
5304 SmallPtrSetImpl<Value *> &AllowedExit) {
5305 Inductions[Phi] = ID;
5306 Type *PhiTy = Phi->getType();
5307 const DataLayout &DL = Phi->getModule()->getDataLayout();
5308
5309 // Get the widest type.
5310 if (!PhiTy->isFloatingPointTy()) {
5311 if (!WidestIndTy)
5312 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
5313 else
5314 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
5315 }
5316
5317 // Int inductions are special because we only allow one IV.
5318 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
5319 ID.getConstIntStepValue() &&
5320 ID.getConstIntStepValue()->isOne() &&
5321 isa<Constant>(ID.getStartValue()) &&
5322 cast<Constant>(ID.getStartValue())->isNullValue()) {
5323
5324 // Use the phi node with the widest type as induction. Use the last
5325 // one if there are multiple (no good reason for doing this other
5326 // than it is expedient). We've checked that it begins at zero and
5327 // steps by one, so this is a canonical induction variable.
5328 if (!PrimaryInduction || PhiTy == WidestIndTy)
5329 PrimaryInduction = Phi;
5330 }
5331
5332 // Both the PHI node itself, and the "post-increment" value feeding
5333 // back into the PHI node may have external users.
5334 AllowedExit.insert(Phi);
5335 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
5336
5337 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)
;
5338 return;
5339}
5340
5341bool LoopVectorizationLegality::canVectorizeInstrs() {
5342 BasicBlock *Header = TheLoop->getHeader();
5343
5344 // Look for the attribute signaling the absence of NaNs.
5345 Function &F = *Header->getParent();
5346 HasFunNoNaNAttr =
5347 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
5348
5349 // For each block in the loop.
5350 for (BasicBlock *BB : TheLoop->blocks()) {
5351 // Scan the instructions in the block and look for hazards.
5352 for (Instruction &I : *BB) {
5353 if (auto *Phi = dyn_cast<PHINode>(&I)) {
5354 Type *PhiTy = Phi->getType();
5355 // Check that this PHI type is allowed.
5356 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
5357 !PhiTy->isPointerTy()) {
5358 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5359 << "loop control flow is not understood by vectorizer");
5360 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)
;
5361 return false;
5362 }
5363
5364 // If this PHINode is not in the header block, then we know that we
5365 // can convert it to select during if-conversion. No need to check if
5366 // the PHIs in this block are induction or reduction variables.
5367 if (BB != Header) {
5368 // Check that this instruction has no outside users or is an
5369 // identified reduction value with an outside user.
5370 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
5371 continue;
5372 ORE->emit(createMissedAnalysis("NeitherInductionNorReduction", Phi)
5373 << "value could not be identified as "
5374 "an induction or reduction variable");
5375 return false;
5376 }
5377
5378 // We only allow if-converted PHIs with exactly two incoming values.
5379 if (Phi->getNumIncomingValues() != 2) {
5380 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5381 << "control flow not understood by vectorizer");
5382 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)
;
5383 return false;
5384 }
5385
5386 RecurrenceDescriptor RedDes;
5387 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
5388 if (RedDes.hasUnsafeAlgebra())
5389 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
5390 AllowedExit.insert(RedDes.getLoopExitInstr());
5391 Reductions[Phi] = RedDes;
5392 continue;
5393 }
5394
5395 InductionDescriptor ID;
5396 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
5397 addInductionPhi(Phi, ID, AllowedExit);
5398 if (ID.hasUnsafeAlgebra() && !HasFunNoNaNAttr)
5399 Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
5400 continue;
5401 }
5402
5403 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
5404 FirstOrderRecurrences.insert(Phi);
5405 continue;
5406 }
5407
5408 // As a last resort, coerce the PHI to a AddRec expression
5409 // and re-try classifying it a an induction PHI.
5410 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
5411 addInductionPhi(Phi, ID, AllowedExit);
5412 continue;
5413 }
5414
5415 ORE->emit(createMissedAnalysis("NonReductionValueUsedOutsideLoop", Phi)
5416 << "value that could not be identified as "
5417 "reduction is used outside the loop");
5418 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)
;
5419 return false;
5420 } // end of PHI handling
5421
5422 // We handle calls that:
5423 // * Are debug info intrinsics.
5424 // * Have a mapping to an IR intrinsic.
5425 // * Have a vector version available.
5426 auto *CI = dyn_cast<CallInst>(&I);
5427 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
5428 !isa<DbgInfoIntrinsic>(CI) &&
5429 !(CI->getCalledFunction() && TLI &&
5430 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
5431 ORE->emit(createMissedAnalysis("CantVectorizeCall", CI)
5432 << "call instruction cannot be vectorized");
5433 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)
;
5434 return false;
5435 }
5436
5437 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
5438 // second argument is the same (i.e. loop invariant)
5439 if (CI && hasVectorInstrinsicScalarOpd(
5440 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
5441 auto *SE = PSE.getSE();
5442 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
5443 ORE->emit(createMissedAnalysis("CantVectorizeIntrinsic", CI)
5444 << "intrinsic instruction cannot be vectorized");
5445 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)
;
5446 return false;
5447 }
5448 }
5449
5450 // Check that the instruction return type is vectorizable.
5451 // Also, we can't vectorize extractelement instructions.
5452 if ((!VectorType::isValidElementType(I.getType()) &&
5453 !I.getType()->isVoidTy()) ||
5454 isa<ExtractElementInst>(I)) {
5455 ORE->emit(createMissedAnalysis("CantVectorizeInstructionReturnType", &I)
5456 << "instruction return type cannot be vectorized");
5457 DEBUG(dbgs() << "LV: Found unvectorizable type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found unvectorizable type.\n"
; } } while (false)
;
5458 return false;
5459 }
5460
5461 // Check that the stored type is vectorizable.
5462 if (auto *ST = dyn_cast<StoreInst>(&I)) {
5463 Type *T = ST->getValueOperand()->getType();
5464 if (!VectorType::isValidElementType(T)) {
5465 ORE->emit(createMissedAnalysis("CantVectorizeStore", ST)
5466 << "store instruction cannot be vectorized");
5467 return false;
5468 }
5469
5470 // FP instructions can allow unsafe algebra, thus vectorizable by
5471 // non-IEEE-754 compliant SIMD units.
5472 // This applies to floating-point math operations and calls, not memory
5473 // operations, shuffles, or casts, as they don't change precision or
5474 // semantics.
5475 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
5476 !I.hasUnsafeAlgebra()) {
5477 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)
;
5478 Hints->setPotentiallyUnsafe();
5479 }
5480
5481 // Reduction instructions are allowed to have exit users.
5482 // All other instructions must not have external users.
5483 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
5484 ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)
5485 << "value cannot be used outside the loop");
5486 return false;
5487 }
5488
5489 } // next instr.
5490 }
5491
5492 if (!PrimaryInduction) {
5493 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)
;
5494 if (Inductions.empty()) {
5495 ORE->emit(createMissedAnalysis("NoInductionVariable")
5496 << "loop induction variable could not be identified");
5497 return false;
5498 }
5499 }
5500
5501 // Now we know the widest induction type, check if our found induction
5502 // is the same size. If it's not, unset it here and InnerLoopVectorizer
5503 // will create another.
5504 if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType())
5505 PrimaryInduction = nullptr;
5506
5507 return true;
5508}
5509
5510void LoopVectorizationCostModel::collectLoopScalars(unsigned VF) {
5511
5512 // We should not collect Scalars more than once per VF. Right now, this
5513 // function is called from collectUniformsAndScalars(), which already does
5514 // this check. Collecting Scalars for VF=1 does not make any sense.
5515 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5516, __PRETTY_FUNCTION__))
5516 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5516, __PRETTY_FUNCTION__))
;
5517
5518 SmallSetVector<Instruction *, 8> Worklist;
5519
5520 // These sets are used to seed the analysis with pointers used by memory
5521 // accesses that will remain scalar.
5522 SmallSetVector<Instruction *, 8> ScalarPtrs;
5523 SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs;
5524
5525 // A helper that returns true if the use of Ptr by MemAccess will be scalar.
5526 // The pointer operands of loads and stores will be scalar as long as the
5527 // memory access is not a gather or scatter operation. The value operand of a
5528 // store will remain scalar if the store is scalarized.
5529 auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) {
5530 InstWidening WideningDecision = getWideningDecision(MemAccess, VF);
5531 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5532, __PRETTY_FUNCTION__))
5532 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5532, __PRETTY_FUNCTION__))
;
5533 if (auto *Store = dyn_cast<StoreInst>(MemAccess))
5534 if (Ptr == Store->getValueOperand())
5535 return WideningDecision == CM_Scalarize;
5536 assert(Ptr == getPointerOperand(MemAccess) &&((Ptr == getPointerOperand(MemAccess) && "Ptr is neither a value or pointer operand"
) ? static_cast<void> (0) : __assert_fail ("Ptr == getPointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5537, __PRETTY_FUNCTION__))
5537 "Ptr is neither a value or pointer operand")((Ptr == getPointerOperand(MemAccess) && "Ptr is neither a value or pointer operand"
) ? static_cast<void> (0) : __assert_fail ("Ptr == getPointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5537, __PRETTY_FUNCTION__))
;
5538 return WideningDecision != CM_GatherScatter;
5539 };
5540
5541 // A helper that returns true if the given value is a bitcast or
5542 // getelementptr instruction contained in the loop.
5543 auto isLoopVaryingBitCastOrGEP = [&](Value *V) {
5544 return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) ||
5545 isa<GetElementPtrInst>(V)) &&
5546 !TheLoop->isLoopInvariant(V);
5547 };
5548
5549 // A helper that evaluates a memory access's use of a pointer. If the use
5550 // will be a scalar use, and the pointer is only used by memory accesses, we
5551 // place the pointer in ScalarPtrs. Otherwise, the pointer is placed in
5552 // PossibleNonScalarPtrs.
5553 auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) {
5554
5555 // We only care about bitcast and getelementptr instructions contained in
5556 // the loop.
5557 if (!isLoopVaryingBitCastOrGEP(Ptr))
5558 return;
5559
5560 // If the pointer has already been identified as scalar (e.g., if it was
5561 // also identified as uniform), there's nothing to do.
5562 auto *I = cast<Instruction>(Ptr);
5563 if (Worklist.count(I))
5564 return;
5565
5566 // If the use of the pointer will be a scalar use, and all users of the
5567 // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise,
5568 // place the pointer in PossibleNonScalarPtrs.
5569 if (isScalarUse(MemAccess, Ptr) && all_of(I->users(), [&](User *U) {
5570 return isa<LoadInst>(U) || isa<StoreInst>(U);
5571 }))
5572 ScalarPtrs.insert(I);
5573 else
5574 PossibleNonScalarPtrs.insert(I);
5575 };
5576
5577 // We seed the scalars analysis with three classes of instructions: (1)
5578 // instructions marked uniform-after-vectorization, (2) bitcast and
5579 // getelementptr instructions used by memory accesses requiring a scalar use,
5580 // and (3) pointer induction variables and their update instructions (we
5581 // currently only scalarize these).
5582 //
5583 // (1) Add to the worklist all instructions that have been identified as
5584 // uniform-after-vectorization.
5585 Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end());
5586
5587 // (2) Add to the worklist all bitcast and getelementptr instructions used by
5588 // memory accesses requiring a scalar use. The pointer operands of loads and
5589 // stores will be scalar as long as the memory accesses is not a gather or
5590 // scatter operation. The value operand of a store will remain scalar if the
5591 // store is scalarized.
5592 for (auto *BB : TheLoop->blocks())
5593 for (auto &I : *BB) {
5594 if (auto *Load = dyn_cast<LoadInst>(&I)) {
5595 evaluatePtrUse(Load, Load->getPointerOperand());
5596 } else if (auto *Store = dyn_cast<StoreInst>(&I)) {
5597 evaluatePtrUse(Store, Store->getPointerOperand());
5598 evaluatePtrUse(Store, Store->getValueOperand());
5599 }
5600 }
5601 for (auto *I : ScalarPtrs)
5602 if (!PossibleNonScalarPtrs.count(I)) {
5603 DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: "
<< *I << "\n"; } } while (false)
;
5604 Worklist.insert(I);
5605 }
5606
5607 // (3) Add to the worklist all pointer induction variables and their update
5608 // instructions.
5609 //
5610 // TODO: Once we are able to vectorize pointer induction variables we should
5611 // no longer insert them into the worklist here.
5612 auto *Latch = TheLoop->getLoopLatch();
5613 for (auto &Induction : *Legal->getInductionVars()) {
5614 auto *Ind = Induction.first;
5615 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5616 if (Induction.second.getKind() != InductionDescriptor::IK_PtrInduction)
5617 continue;
5618 Worklist.insert(Ind);
5619 Worklist.insert(IndUpdate);
5620 DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: "
<< *Ind << "\n"; } } while (false)
;
5621 DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: "
<< *IndUpdate << "\n"; } } while (false)
;
5622 }
5623
5624 // Expand the worklist by looking through any bitcasts and getelementptr
5625 // instructions we've already identified as scalar. This is similar to the
5626 // expansion step in collectLoopUniforms(); however, here we're only
5627 // expanding to include additional bitcasts and getelementptr instructions.
5628 unsigned Idx = 0;
5629 while (Idx != Worklist.size()) {
5630 Instruction *Dst = Worklist[Idx++];
5631 if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0)))
5632 continue;
5633 auto *Src = cast<Instruction>(Dst->getOperand(0));
5634 if (all_of(Src->users(), [&](User *U) -> bool {
5635 auto *J = cast<Instruction>(U);
5636 return !TheLoop->contains(J) || Worklist.count(J) ||
5637 ((isa<LoadInst>(J) || isa<StoreInst>(J)) &&
5638 isScalarUse(J, Src));
5639 })) {
5640 Worklist.insert(Src);
5641 DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: "
<< *Src << "\n"; } } while (false)
;
5642 }
5643 }
5644
5645 // An induction variable will remain scalar if all users of the induction
5646 // variable and induction variable update remain scalar.
5647 for (auto &Induction : *Legal->getInductionVars()) {
5648 auto *Ind = Induction.first;
5649 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5650
5651 // We already considered pointer induction variables, so there's no reason
5652 // to look at their users again.
5653 //
5654 // TODO: Once we are able to vectorize pointer induction variables we
5655 // should no longer skip over them here.
5656 if (Induction.second.getKind() == InductionDescriptor::IK_PtrInduction)
5657 continue;
5658
5659 // Determine if all users of the induction variable are scalar after
5660 // vectorization.
5661 auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
5662 auto *I = cast<Instruction>(U);
5663 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I);
5664 });
5665 if (!ScalarInd)
5666 continue;
5667
5668 // Determine if all users of the induction variable update instruction are
5669 // scalar after vectorization.
5670 auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5671 auto *I = cast<Instruction>(U);
5672 return I == Ind || !TheLoop->contains(I) || Worklist.count(I);
5673 });
5674 if (!ScalarIndUpdate)
5675 continue;
5676
5677 // The induction variable and its update instruction will remain scalar.
5678 Worklist.insert(Ind);
5679 Worklist.insert(IndUpdate);
5680 DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: "
<< *Ind << "\n"; } } while (false)
;
5681 DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: "
<< *IndUpdate << "\n"; } } while (false)
;
5682 }
5683
5684 Scalars[VF].insert(Worklist.begin(), Worklist.end());
5685}
5686
5687bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
5688 if (!blockNeedsPredication(I->getParent()))
5689 return false;
5690 switch(I->getOpcode()) {
5691 default:
5692 break;
5693 case Instruction::Store:
5694 return !isMaskRequired(I);
5695 case Instruction::UDiv:
5696 case Instruction::SDiv:
5697 case Instruction::SRem:
5698 case Instruction::URem:
5699 return mayDivideByZero(*I);
5700 }
5701 return false;
5702}
5703
5704bool LoopVectorizationLegality::memoryInstructionCanBeWidened(Instruction *I,
5705 unsigned VF) {
5706 // Get and ensure we have a valid memory instruction.
5707 LoadInst *LI = dyn_cast<LoadInst>(I);
5708 StoreInst *SI = dyn_cast<StoreInst>(I);
5709 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5709, __PRETTY_FUNCTION__))
;
5710
5711 auto *Ptr = getPointerOperand(I);
5712
5713 // In order to be widened, the pointer should be consecutive, first of all.
5714 if (!isConsecutivePtr(Ptr))
5715 return false;
5716
5717 // If the instruction is a store located in a predicated block, it will be
5718 // scalarized.
5719 if (isScalarWithPredication(I))
5720 return false;
5721
5722 // If the instruction's allocated size doesn't equal it's type size, it
5723 // requires padding and will be scalarized.
5724 auto &DL = I->getModule()->getDataLayout();
5725 auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
5726 if (hasIrregularType(ScalarTy, DL, VF))
5727 return false;
5728
5729 return true;
5730}
5731
5732void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) {
5733
5734 // We should not collect Uniforms more than once per VF. Right now,
5735 // this function is called from collectUniformsAndScalars(), which
5736 // already does this check. Collecting Uniforms for VF=1 does not make any
5737 // sense.
5738
5739 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5740, __PRETTY_FUNCTION__))
5740 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5740, __PRETTY_FUNCTION__))
;
5741
5742 // Visit the list of Uniforms. If we'll not find any uniform value, we'll
5743 // not analyze again. Uniforms.count(VF) will return 1.
5744 Uniforms[VF].clear();
5745
5746 // We now know that the loop is vectorizable!
5747 // Collect instructions inside the loop that will remain uniform after
5748 // vectorization.
5749
5750 // Global values, params and instructions outside of current loop are out of
5751 // scope.
5752 auto isOutOfScope = [&](Value *V) -> bool {
5753 Instruction *I = dyn_cast<Instruction>(V);
5754 return (!I || !TheLoop->contains(I));
5755 };
5756
5757 SetVector<Instruction *> Worklist;
5758 BasicBlock *Latch = TheLoop->getLoopLatch();
5759
5760 // Start with the conditional branch. If the branch condition is an
5761 // instruction contained in the loop that is only used by the branch, it is
5762 // uniform.
5763 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
5764 if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) {
5765 Worklist.insert(Cmp);
5766 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)
;
5767 }
5768
5769 // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
5770 // are pointers that are treated like consecutive pointers during
5771 // vectorization. The pointer operands of interleaved accesses are an
5772 // example.
5773 SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
5774
5775 // Holds pointer operands of instructions that are possibly non-uniform.
5776 SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
5777
5778 auto isUniformDecision = [&](Instruction *I, unsigned VF) {
5779 InstWidening WideningDecision = getWideningDecision(I, VF);
5780 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5781, __PRETTY_FUNCTION__))
5781 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 5781, __PRETTY_FUNCTION__))
;
5782
5783 return (WideningDecision == CM_Widen ||
5784 WideningDecision == CM_Interleave);
5785 };
5786 // Iterate over the instructions in the loop, and collect all
5787 // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
5788 // that a consecutive-like pointer operand will be scalarized, we collect it
5789 // in PossibleNonUniformPtrs instead. We use two sets here because a single
5790 // getelementptr instruction can be used by both vectorized and scalarized
5791 // memory instructions. For example, if a loop loads and stores from the same
5792 // location, but the store is conditional, the store will be scalarized, and
5793 // the getelementptr won't remain uniform.
5794 for (auto *BB : TheLoop->blocks())
5795 for (auto &I : *BB) {
5796
5797 // If there's no pointer operand, there's nothing to do.
5798 auto *Ptr = dyn_cast_or_null<Instruction>(getPointerOperand(&I));
5799 if (!Ptr)
5800 continue;
5801
5802 // True if all users of Ptr are memory accesses that have Ptr as their
5803 // pointer operand.
5804 auto UsersAreMemAccesses = all_of(Ptr->users(), [&](User *U) -> bool {
5805 return getPointerOperand(U) == Ptr;
5806 });
5807
5808 // Ensure the memory instruction will not be scalarized or used by
5809 // gather/scatter, making its pointer operand non-uniform. If the pointer
5810 // operand is used by any instruction other than a memory access, we
5811 // conservatively assume the pointer operand may be non-uniform.
5812 if (!UsersAreMemAccesses || !isUniformDecision(&I, VF))
5813 PossibleNonUniformPtrs.insert(Ptr);
5814
5815 // If the memory instruction will be vectorized and its pointer operand
5816 // is consecutive-like, or interleaving - the pointer operand should
5817 // remain uniform.
5818 else
5819 ConsecutiveLikePtrs.insert(Ptr);
5820 }
5821
5822 // Add to the Worklist all consecutive and consecutive-like pointers that
5823 // aren't also identified as possibly non-uniform.
5824 for (auto *V : ConsecutiveLikePtrs)
5825 if (!PossibleNonUniformPtrs.count(V)) {
5826 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)
;
5827 Worklist.insert(V);
5828 }
5829
5830 // Expand Worklist in topological order: whenever a new instruction
5831 // is added , its users should be either already inside Worklist, or
5832 // out of scope. It ensures a uniform instruction will only be used
5833 // by uniform instructions or out of scope instructions.
5834 unsigned idx = 0;
5835 while (idx != Worklist.size()) {
5836 Instruction *I = Worklist[idx++];
5837
5838 for (auto OV : I->operand_values()) {
5839 if (isOutOfScope(OV))
5840 continue;
5841 auto *OI = cast<Instruction>(OV);
5842 if (all_of(OI->users(), [&](User *U) -> bool {
5843 auto *J = cast<Instruction>(U);
5844 return !TheLoop->contains(J) || Worklist.count(J) ||
5845 (OI == getPointerOperand(J) && isUniformDecision(J, VF));
5846 })) {
5847 Worklist.insert(OI);
5848 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)
;
5849 }
5850 }
5851 }
5852
5853 // Returns true if Ptr is the pointer operand of a memory access instruction
5854 // I, and I is known to not require scalarization.
5855 auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
5856 return getPointerOperand(I) == Ptr && isUniformDecision(I, VF);
5857 };
5858
5859 // For an instruction to be added into Worklist above, all its users inside
5860 // the loop should also be in Worklist. However, this condition cannot be
5861 // true for phi nodes that form a cyclic dependence. We must process phi
5862 // nodes separately. An induction variable will remain uniform if all users
5863 // of the induction variable and induction variable update remain uniform.
5864 // The code below handles both pointer and non-pointer induction variables.
5865 for (auto &Induction : *Legal->getInductionVars()) {
5866 auto *Ind = Induction.first;
5867 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5868
5869 // Determine if all users of the induction variable are uniform after
5870 // vectorization.
5871 auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
5872 auto *I = cast<Instruction>(U);
5873 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
5874 isVectorizedMemAccessUse(I, Ind);
5875 });
5876 if (!UniformInd)
5877 continue;
5878
5879 // Determine if all users of the induction variable update instruction are
5880 // uniform after vectorization.
5881 auto UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5882 auto *I = cast<Instruction>(U);
5883 return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||
5884 isVectorizedMemAccessUse(I, IndUpdate);
5885 });
5886 if (!UniformIndUpdate)
5887 continue;
5888
5889 // The induction variable and its update instruction will remain uniform.
5890 Worklist.insert(Ind);
5891 Worklist.insert(IndUpdate);
5892 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)
;
5893 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)
;
5894 }
5895
5896 Uniforms[VF].insert(Worklist.begin(), Worklist.end());
5897}
5898
5899bool LoopVectorizationLegality::canVectorizeMemory() {
5900 LAI = &(*GetLAA)(*TheLoop);
5901 InterleaveInfo.setLAI(LAI);
5902 const OptimizationRemarkAnalysis *LAR = LAI->getReport();
5903 if (LAR) {
5904 OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
5905 "loop not vectorized: ", *LAR);
5906 ORE->emit(VR);
5907 }
5908 if (!LAI->canVectorizeMemory())
5909 return false;
5910
5911 if (LAI->hasStoreToLoopInvariantAddress()) {
5912 ORE->emit(createMissedAnalysis("CantVectorizeStoreToLoopInvariantAddress")
5913 << "write to a loop invariant address could not be vectorized");
5914 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)
;
5915 return false;
5916 }
5917
5918 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
5919 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
5920
5921 return true;
5922}
5923
5924bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5925 Value *In0 = const_cast<Value *>(V);
5926 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5927 if (!PN)
5928 return false;
5929
5930 return Inductions.count(PN);
5931}
5932
5933bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
5934 return FirstOrderRecurrences.count(Phi);
5935}
5936
5937bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5938 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
5939}
5940
5941bool LoopVectorizationLegality::blockCanBePredicated(
5942 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
5943 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
5944
5945 for (Instruction &I : *BB) {
5946 // Check that we don't have a constant expression that can trap as operand.
5947 for (Value *Operand : I.operands()) {
5948 if (auto *C = dyn_cast<Constant>(Operand))
5949 if (C->canTrap())
5950 return false;
5951 }
5952 // We might be able to hoist the load.
5953 if (I.mayReadFromMemory()) {
5954 auto *LI = dyn_cast<LoadInst>(&I);
5955 if (!LI)
5956 return false;
5957 if (!SafePtrs.count(LI->getPointerOperand())) {
5958 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
5959 isLegalMaskedGather(LI->getType())) {
5960 MaskedOp.insert(LI);
5961 continue;
5962 }
5963 // !llvm.mem.parallel_loop_access implies if-conversion safety.
5964 if (IsAnnotatedParallel)
5965 continue;
5966 return false;
5967 }
5968 }
5969
5970 if (I.mayWriteToMemory()) {
5971 auto *SI = dyn_cast<StoreInst>(&I);
5972 // We only support predication of stores in basic blocks with one
5973 // predecessor.
5974 if (!SI)
5975 return false;
5976
5977 // Build a masked store if it is legal for the target.
5978 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
5979 SI->getPointerOperand()) ||
5980 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
5981 MaskedOp.insert(SI);
5982 continue;
5983 }
5984
5985 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5986 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5987
5988 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5989 !isSinglePredecessor)
5990 return false;
5991 }
5992 if (I.mayThrow())
5993 return false;
5994 }
5995
5996 return true;
5997}
5998
5999void InterleavedAccessInfo::collectConstStrideAccesses(
6000 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
6001 const ValueToValueMap &Strides) {
6002
6003 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
6004
6005 // Since it's desired that the load/store instructions be maintained in
6006 // "program order" for the interleaved access analysis, we have to visit the
6007 // blocks in the loop in reverse postorder (i.e., in a topological order).
6008 // Such an ordering will ensure that any load/store that may be executed
6009 // before a second load/store will precede the second load/store in
6010 // AccessStrideInfo.
6011 LoopBlocksDFS DFS(TheLoop);
6012 DFS.perform(LI);
6013 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
6014 for (auto &I : *BB) {
6015 auto *LI = dyn_cast<LoadInst>(&I);
6016 auto *SI = dyn_cast<StoreInst>(&I);
6017 if (!LI && !SI)
6018 continue;
6019
6020 Value *Ptr = getPointerOperand(&I);
6021 // We don't check wrapping here because we don't know yet if Ptr will be
6022 // part of a full group or a group with gaps. Checking wrapping for all
6023 // pointers (even those that end up in groups with no gaps) will be overly
6024 // conservative. For full groups, wrapping should be ok since if we would
6025 // wrap around the address space we would do a memory access at nullptr
6026 // even without the transformation. The wrapping checks are therefore
6027 // deferred until after we've formed the interleaved groups.
6028 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
6029 /*Assume=*/true, /*ShouldCheckWrap=*/false);
6030
6031 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
6032 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
6033 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
6034
6035 // An alignment of 0 means target ABI alignment.
6036 unsigned Align = getMemInstAlignment(&I);
6037 if (!Align)
6038 Align = DL.getABITypeAlignment(PtrTy->getElementType());
6039
6040 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
6041 }
6042}
6043
6044// Analyze interleaved accesses and collect them into interleaved load and
6045// store groups.
6046//
6047// When generating code for an interleaved load group, we effectively hoist all
6048// loads in the group to the location of the first load in program order. When
6049// generating code for an interleaved store group, we sink all stores to the
6050// location of the last store. This code motion can change the order of load
6051// and store instructions and may break dependences.
6052//
6053// The code generation strategy mentioned above ensures that we won't violate
6054// any write-after-read (WAR) dependences.
6055//
6056// E.g., for the WAR dependence: a = A[i]; // (1)
6057// A[i] = b; // (2)
6058//
6059// The store group of (2) is always inserted at or below (2), and the load
6060// group of (1) is always inserted at or above (1). Thus, the instructions will
6061// never be reordered. All other dependences are checked to ensure the
6062// correctness of the instruction reordering.
6063//
6064// The algorithm visits all memory accesses in the loop in bottom-up program
6065// order. Program order is established by traversing the blocks in the loop in
6066// reverse postorder when collecting the accesses.
6067//
6068// We visit the memory accesses in bottom-up order because it can simplify the
6069// construction of store groups in the presence of write-after-write (WAW)
6070// dependences.
6071//
6072// E.g., for the WAW dependence: A[i] = a; // (1)
6073// A[i] = b; // (2)
6074// A[i + 1] = c; // (3)
6075//
6076// We will first create a store group with (3) and (2). (1) can't be added to
6077// this group because it and (2) are dependent. However, (1) can be grouped
6078// with other accesses that may precede it in program order. Note that a
6079// bottom-up order does not imply that WAW dependences should not be checked.
6080void InterleavedAccessInfo::analyzeInterleaving(
6081 const ValueToValueMap &Strides) {
6082 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Analyzing interleaved accesses...\n"
; } } while (false)
;
6083
6084 // Holds all accesses with a constant stride.
6085 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
6086 collectConstStrideAccesses(AccessStrideInfo, Strides);
6087
6088 if (AccessStrideInfo.empty())
6089 return;
6090
6091 // Collect the dependences in the loop.
6092 collectDependences();
6093
6094 // Holds all interleaved store groups temporarily.
6095 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
6096 // Holds all interleaved load groups temporarily.
6097 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
6098
6099 // Search in bottom-up program order for pairs of accesses (A and B) that can
6100 // form interleaved load or store groups. In the algorithm below, access A
6101 // precedes access B in program order. We initialize a group for B in the
6102 // outer loop of the algorithm, and then in the inner loop, we attempt to
6103 // insert each A into B's group if:
6104 //
6105 // 1. A and B have the same stride,
6106 // 2. A and B have the same memory object size, and
6107 // 3. A belongs in B's group according to its distance from B.
6108 //
6109 // Special care is taken to ensure group formation will not break any
6110 // dependences.
6111 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
6112 BI != E; ++BI) {
6113 Instruction *B = BI->first;
6114 StrideDescriptor DesB = BI->second;
6115
6116 // Initialize a group for B if it has an allowable stride. Even if we don't
6117 // create a group for B, we continue with the bottom-up algorithm to ensure
6118 // we don't break any of B's dependences.
6119 InterleaveGroup *Group = nullptr;
6120 if (isStrided(DesB.Stride)) {
6121 Group = getInterleaveGroup(B);
6122 if (!Group) {
6123 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)
;
6124 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
6125 }
6126 if (B->mayWriteToMemory())
6127 StoreGroups.insert(Group);
6128 else
6129 LoadGroups.insert(Group);
6130 }
6131
6132 for (auto AI = std::next(BI); AI != E; ++AI) {
6133 Instruction *A = AI->first;
6134 StrideDescriptor DesA = AI->second;
6135
6136 // Our code motion strategy implies that we can't have dependences
6137 // between accesses in an interleaved group and other accesses located
6138 // between the first and last member of the group. Note that this also
6139 // means that a group can't have more than one member at a given offset.
6140 // The accesses in a group can have dependences with other accesses, but
6141 // we must ensure we don't extend the boundaries of the group such that
6142 // we encompass those dependent accesses.
6143 //
6144 // For example, assume we have the sequence of accesses shown below in a
6145 // stride-2 loop:
6146 //
6147 // (1, 2) is a group | A[i] = a; // (1)
6148 // | A[i-1] = b; // (2) |
6149 // A[i-3] = c; // (3)
6150 // A[i] = d; // (4) | (2, 4) is not a group
6151 //
6152 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
6153 // but not with (4). If we did, the dependent access (3) would be within
6154 // the boundaries of the (2, 4) group.
6155 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
6156
6157 // If a dependence exists and A is already in a group, we know that A
6158 // must be a store since A precedes B and WAR dependences are allowed.
6159 // Thus, A would be sunk below B. We release A's group to prevent this
6160 // illegal code motion. A will then be free to form another group with
6161 // instructions that precede it.
6162 if (isInterleaved(A)) {
6163 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
6164 StoreGroups.remove(StoreGroup);
6165 releaseGroup(StoreGroup);
6166 }
6167
6168 // If a dependence exists and A is not already in a group (or it was
6169 // and we just released it), B might be hoisted above A (if B is a
6170 // load) or another store might be sunk below A (if B is a store). In
6171 // either case, we can't add additional instructions to B's group. B
6172 // will only form a group with instructions that it precedes.
6173 break;
6174 }
6175
6176 // At this point, we've checked for illegal code motion. If either A or B
6177 // isn't strided, there's nothing left to do.
6178 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
6179 continue;
6180
6181 // Ignore A if it's already in a group or isn't the same kind of memory
6182 // operation as B.
6183 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
6184 continue;
6185
6186 // Check rules 1 and 2. Ignore A if its stride or size is different from
6187 // that of B.
6188 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
6189 continue;
6190
6191 // Ignore A if the memory object of A and B don't belong to the same
6192 // address space
6193 if (getMemInstAddressSpace(A) != getMemInstAddressSpace(B))
6194 continue;
6195
6196 // Calculate the distance from A to B.
6197 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
6198 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
6199 if (!DistToB)
6200 continue;
6201 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
6202
6203 // Check rule 3. Ignore A if its distance to B is not a multiple of the
6204 // size.
6205 if (DistanceToB % static_cast<int64_t>(DesB.Size))
6206 continue;
6207
6208 // Ignore A if either A or B is in a predicated block. Although we
6209 // currently prevent group formation for predicated accesses, we may be
6210 // able to relax this limitation in the future once we handle more
6211 // complicated blocks.
6212 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
6213 continue;
6214
6215 // The index of A is the index of B plus A's distance to B in multiples
6216 // of the size.
6217 int IndexA =
6218 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
6219
6220 // Try to insert A into B's group.
6221 if (Group->insertMember(A, IndexA, DesA.Align)) {
6222 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)
6223 << " 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)
;
6224 InterleaveGroupMap[A] = Group;
6225
6226 // Set the first load in program order as the insert position.
6227 if (A->mayReadFromMemory())
6228 Group->setInsertPos(A);
6229 }
6230 } // Iteration over A accesses.
6231 } // Iteration over B accesses.
6232
6233 // Remove interleaved store groups with gaps.
6234 for (InterleaveGroup *Group : StoreGroups)
6235 if (Group->getNumMembers() != Group->getFactor())
6236 releaseGroup(Group);
6237
6238 // Remove interleaved groups with gaps (currently only loads) whose memory
6239 // accesses may wrap around. We have to revisit the getPtrStride analysis,
6240 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
6241 // not check wrapping (see documentation there).
6242 // FORNOW we use Assume=false;
6243 // TODO: Change to Assume=true but making sure we don't exceed the threshold
6244 // of runtime SCEV assumptions checks (thereby potentially failing to
6245 // vectorize altogether).
6246 // Additional optional optimizations:
6247 // TODO: If we are peeling the loop and we know that the first pointer doesn't
6248 // wrap then we can deduce that all pointers in the group don't wrap.
6249 // This means that we can forcefully peel the loop in order to only have to
6250 // check the first pointer for no-wrap. When we'll change to use Assume=true
6251 // we'll only need at most one runtime check per interleaved group.
6252 //
6253 for (InterleaveGroup *Group : LoadGroups) {
6254
6255 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
6256 // load would wrap around the address space we would do a memory access at
6257 // nullptr even without the transformation.
6258 if (Group->getNumMembers() == Group->getFactor())
6259 continue;
6260
6261 // Case 2: If first and last members of the group don't wrap this implies
6262 // that all the pointers in the group don't wrap.
6263 // So we check only group member 0 (which is always guaranteed to exist),
6264 // and group member Factor - 1; If the latter doesn't exist we rely on
6265 // peeling (if it is a non-reveresed accsess -- see Case 3).
6266 Value *FirstMemberPtr = getPointerOperand(Group->getMember(0));
6267 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
6268 /*ShouldCheckWrap=*/true)) {
6269 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)
6270 "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)
;
6271 releaseGroup(Group);
6272 continue;
6273 }
6274 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
6275 if (LastMember) {
6276 Value *LastMemberPtr = getPointerOperand(LastMember);
6277 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
6278 /*ShouldCheckWrap=*/true)) {
6279 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)
6280 "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)
;
6281 releaseGroup(Group);
6282 }
6283 } else {
6284 // Case 3: A non-reversed interleaved load group with gaps: We need
6285 // to execute at least one scalar epilogue iteration. This will ensure
6286 // we don't speculatively access memory out-of-bounds. We only need
6287 // to look for a member at index factor - 1, since every group must have
6288 // a member at index zero.
6289 if (Group->isReverse()) {
6290 releaseGroup(Group);
6291 continue;
6292 }
6293 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)
;
6294 RequiresScalarEpilogue = true;
6295 }
6296 }
6297}
6298
6299Optional<unsigned> LoopVectorizationCostModel::computeMaxVF(bool OptForSize) {
6300 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
6301 ORE->emit(createMissedAnalysis("ConditionalStore")
6302 << "store that is conditionally executed prevents vectorization");
6303 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)
;
6304 return None;
6305 }
6306
6307 if (!OptForSize) // Remaining checks deal with scalar loop when OptForSize.
6308 return computeFeasibleMaxVF(OptForSize);
6309
6310 if (Legal->getRuntimePointerChecking()->Need) {
6311 ORE->emit(createMissedAnalysis("CantVersionLoopWithOptForSize")
6312 << "runtime pointer checks needed. Enable vectorization of this "
6313 "loop with '#pragma clang loop vectorize(enable)' when "
6314 "compiling with -Os/-Oz");
6315 DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n"
; } } while (false)
6316 << "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)
;
6317 return None;
6318 }
6319
6320 // If we optimize the program for size, avoid creating the tail loop.
6321 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6322 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)
;
6323
6324 // If we don't know the precise trip count, don't try to vectorize.
6325 if (TC < 2) {
6326 ORE->emit(
6327 createMissedAnalysis("UnknownLoopCountComplexCFG")
6328 << "unable to calculate the loop count due to complex control flow");
6329 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)
;
6330 return None;
6331 }
6332
6333 unsigned MaxVF = computeFeasibleMaxVF(OptForSize);
6334
6335 if (TC % MaxVF != 0) {
6336 // If the trip count that we found modulo the vectorization factor is not
6337 // zero then we require a tail.
6338 // FIXME: look for a smaller MaxVF that does divide TC rather than give up.
6339 // FIXME: return None if loop requiresScalarEpilog(<MaxVF>), or look for a
6340 // smaller MaxVF that does not require a scalar epilog.
6341
6342 ORE->emit(createMissedAnalysis("NoTailLoopWithOptForSize")
6343 << "cannot optimize for size and vectorize at the "
6344 "same time. Enable vectorization of this loop "
6345 "with '#pragma clang loop vectorize(enable)' "
6346 "when compiling with -Os/-Oz");
6347 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)
;
6348 return None;
6349 }
6350
6351 return MaxVF;
6352}
6353
6354unsigned LoopVectorizationCostModel::computeFeasibleMaxVF(bool OptForSize) {
6355 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
6356 unsigned SmallestType, WidestType;
6357 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
6358 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
6359 unsigned MaxSafeDepDist = -1U;
6360
6361 // Get the maximum safe dependence distance in bits computed by LAA. If the
6362 // loop contains any interleaved accesses, we divide the dependence distance
6363 // by the maximum interleave factor of all interleaved groups. Note that
6364 // although the division ensures correctness, this is a fairly conservative
6365 // computation because the maximum distance computed by LAA may not involve
6366 // any of the interleaved accesses.
6367 if (Legal->getMaxSafeDepDistBytes() != -1U)
6368 MaxSafeDepDist =
6369 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
6370
6371 WidestRegister =
6372 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
6373 unsigned MaxVectorSize = WidestRegister / WidestType;
6374
6375 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)
6376 << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: "
<< SmallestType << " / " << WidestType <<
" bits.\n"; } } while (false)
;
6377 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
)
6378 << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The Widest register is: "
<< WidestRegister << " bits.\n"; } } while (false
)
;
6379
6380 if (MaxVectorSize == 0) {
6381 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)
;
6382 MaxVectorSize = 1;
6383 }
6384
6385 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6386, __PRETTY_FUNCTION__))
6386 " 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6386, __PRETTY_FUNCTION__))
;
6387
6388 unsigned MaxVF = MaxVectorSize;
6389 if (MaximizeBandwidth && !OptForSize) {
6390 // Collect all viable vectorization factors.
6391 SmallVector<unsigned, 8> VFs;
6392 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
6393 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
6394 VFs.push_back(VS);
6395
6396 // For each VF calculate its register usage.
6397 auto RUs = calculateRegisterUsage(VFs);
6398
6399 // Select the largest VF which doesn't require more registers than existing
6400 // ones.
6401 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
6402 for (int i = RUs.size() - 1; i >= 0; --i) {
6403 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
6404 MaxVF = VFs[i];
6405 break;
6406 }
6407 }
6408 }
6409 return MaxVF;
6410}
6411
6412LoopVectorizationCostModel::VectorizationFactor
6413LoopVectorizationCostModel::selectVectorizationFactor(unsigned MaxVF) {
6414 float Cost = expectedCost(1).first;
6415#ifndef NDEBUG
6416 const float ScalarCost = Cost;
6417#endif /* NDEBUG */
6418 unsigned Width = 1;
6419 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)
;
6420
6421 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
6422 // Ignore scalar width, because the user explicitly wants vectorization.
6423 if (ForceVectorization && MaxVF > 1) {
6424 Width = 2;
6425 Cost = expectedCost(Width).first / (float)Width;
6426 }
6427
6428 for (unsigned i = 2; i <= MaxVF; i *= 2) {
6429 // Notice that the vector loop needs to be executed less times, so
6430 // we need to divide the cost of the vector loops by the width of
6431 // the vector elements.
6432 VectorizationCostTy C = expectedCost(i);
6433 float VectorCost = C.first / (float)i;
6434 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)
6435 << " 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)
;
6436 if (!C.second && !ForceVectorization) {
6437 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)
6438 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)
6439 << " 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)
;
6440 continue;
6441 }
6442 if (VectorCost < Cost) {
6443 Cost = VectorCost;
6444 Width = i;
6445 }
6446 }
6447
6448 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)
6449 << "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)
6450 << "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)
;
6451 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Selecting VF: " <<
Width << ".\n"; } } while (false)
;
6452 VectorizationFactor Factor = {Width, (unsigned)(Width * Cost)};
6453 return Factor;
6454}
6455
6456std::pair<unsigned, unsigned>
6457LoopVectorizationCostModel::getSmallestAndWidestTypes() {
6458 unsigned MinWidth = -1U;
6459 unsigned MaxWidth = 8;
6460 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6461
6462 // For each block.
6463 for (BasicBlock *BB : TheLoop->blocks()) {
6464 // For each instruction in the loop.
6465 for (Instruction &I : *BB) {
6466 Type *T = I.getType();
6467
6468 // Skip ignored values.
6469 if (ValuesToIgnore.count(&I))
6470 continue;
6471
6472 // Only examine Loads, Stores and PHINodes.
6473 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
6474 continue;
6475
6476 // Examine PHI nodes that are reduction variables. Update the type to
6477 // account for the recurrence type.
6478 if (auto *PN = dyn_cast<PHINode>(&I)) {
6479 if (!Legal->isReductionVariable(PN))
6480 continue;
6481 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
6482 T = RdxDesc.getRecurrenceType();
6483 }
6484
6485 // Examine the stored values.
6486 if (auto *ST = dyn_cast<StoreInst>(&I))
6487 T = ST->getValueOperand()->getType();
6488
6489 // Ignore loaded pointer types and stored pointer types that are not
6490 // vectorizable.
6491 //
6492 // FIXME: The check here attempts to predict whether a load or store will
6493 // be vectorized. We only know this for certain after a VF has
6494 // been selected. Here, we assume that if an access can be
6495 // vectorized, it will be. We should also look at extending this
6496 // optimization to non-pointer types.
6497 //
6498 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I) &&
6499 !Legal->isAccessInterleaved(&I) && !Legal->isLegalGatherOrScatter(&I))
6500 continue;
6501
6502 MinWidth = std::min(MinWidth,
6503 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6504 MaxWidth = std::max(MaxWidth,
6505 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6506 }
6507 }
6508
6509 return {MinWidth, MaxWidth};
6510}
6511
6512unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
6513 unsigned VF,
6514 unsigned LoopCost) {
6515
6516 // -- The interleave heuristics --
6517 // We interleave the loop in order to expose ILP and reduce the loop overhead.
6518 // There are many micro-architectural considerations that we can't predict
6519 // at this level. For example, frontend pressure (on decode or fetch) due to
6520 // code size, or the number and capabilities of the execution ports.
6521 //
6522 // We use the following heuristics to select the interleave count:
6523 // 1. If the code has reductions, then we interleave to break the cross
6524 // iteration dependency.
6525 // 2. If the loop is really small, then we interleave to reduce the loop
6526 // overhead.
6527 // 3. We don't interleave if we think that we will spill registers to memory
6528 // due to the increased register pressure.
6529
6530 // When we optimize for size, we don't interleave.
6531 if (OptForSize)
1
Assuming 'OptForSize' is 0
2
Taking false branch
6532 return 1;
6533
6534 // We used the distance for the interleave count.
6535 if (Legal->getMaxSafeDepDistBytes() != -1U)
3
Assuming the condition is false
4
Taking false branch
6536 return 1;
6537
6538 // Do not interleave loops with a relatively small trip count.
6539 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6540 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
5
Assuming 'TC' is <= 1
6541 return 1;
6542
6543 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
6
Assuming 'VF' is <= 1
6544 DEBUG(dbgs() << "LV: The target has " << TargetNumRegistersdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (false
)
6545 << " registers\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: The target has " <<
TargetNumRegisters << " registers\n"; } } while (false
)
;
6546
6547 if (VF == 1) {
7
Assuming 'VF' is not equal to 1
8
Taking false branch
6548 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
6549 TargetNumRegisters = ForceTargetNumScalarRegs;
6550 } else {
6551 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
9
Assuming the condition is false
10
Taking false branch
6552 TargetNumRegisters = ForceTargetNumVectorRegs;
6553 }
6554
6555 RegisterUsage R = calculateRegisterUsage({VF})[0];
6556 // We divide by these constants so assume that we have at least one
6557 // instruction that uses at least one register.
6558 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
6559 R.NumInstructions = std::max(R.NumInstructions, 1U);
6560
6561 // We calculate the interleave count using the following formula.
6562 // Subtract the number of loop invariants from the number of available
6563 // registers. These registers are used by all of the interleaved instances.
6564 // Next, divide the remaining registers by the number of registers that is
6565 // required by the loop, in order to estimate how many parallel instances
6566 // fit without causing spills. All of this is rounded down if necessary to be
6567 // a power of two. We want power of two interleave count to simplify any
6568 // addressing operations or alignment considerations.
6569 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
6570 R.MaxLocalUsers);
6571
6572 // Don't count the induction variable as interleaved.
6573 if (EnableIndVarRegisterHeur)
11
Assuming the condition is false
12
Taking false branch
6574 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
6575 std::max(1U, (R.MaxLocalUsers - 1)));
6576
6577 // Clamp the interleave ranges to reasonable counts.
6578 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
6579
6580 // Check if the user has overridden the max.
6581 if (VF == 1) {
13
Taking false branch
6582 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
6583 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
6584 } else {
6585 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
14
Assuming the condition is false
15
Taking false branch
6586 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
6587 }
6588
6589 // If we did not calculate the cost for VF (because the user selected the VF)
6590 // then we calculate the cost of VF here.
6591 if (LoopCost == 0)
16
Assuming 'LoopCost' is equal to 0
17
Taking true branch
6592 LoopCost = expectedCost(VF).first;
18
The value 0 is assigned to 'LoopCost'
6593
6594 // Clamp the calculated IC to be between the 1 and the max interleave count
6595 // that the target allows.
6596 if (IC > MaxInterleaveCount)
19
Taking false branch
6597 IC = MaxInterleaveCount;
6598 else if (IC < 1)
20
Assuming 'IC' is >= 1
21
Taking false branch
6599 IC = 1;
6600
6601 // Interleave if we vectorized this loop and there is a reduction that could
6602 // benefit from interleaving.
6603 if (VF > 1 && Legal->getReductionVars()->size()) {
6604 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)
;
6605 return IC;
6606 }
6607
6608 // Note that if we've already vectorized the loop we will have done the
6609 // runtime check and so interleaving won't require further checks.
6610 bool InterleavingRequiresRuntimePointerCheck =
6611 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
6612
6613 // We want to interleave small loops in order to reduce the loop overhead and
6614 // potentially expose ILP opportunities.
6615 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)
;
6616 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
22
Assuming the condition is true
23
Taking true branch
6617 // We assume that the cost overhead is 1 and we use the cost model
6618 // to estimate the cost of the loop and interleave until the cost of the
6619 // loop overhead is about 5% of the cost of the loop.
6620 unsigned SmallIC =
6621 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
24
Division by zero
6622
6623 // Interleave until store/load ports (estimated by max interleave count) are
6624 // saturated.
6625 unsigned NumStores = Legal->getNumStores();
6626 unsigned NumLoads = Legal->getNumLoads();
6627 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
6628 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
6629
6630 // If we have a scalar reduction (vector reductions are already dealt with
6631 // by this point), we can increase the critical path length if the loop
6632 // we're interleaving is inside another loop. Limit, by default to 2, so the
6633 // critical path only gets increased by one reduction operation.
6634 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
6635 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
6636 SmallIC = std::min(SmallIC, F);
6637 StoresIC = std::min(StoresIC, F);
6638 LoadsIC = std::min(LoadsIC, F);
6639 }
6640
6641 if (EnableLoadStoreRuntimeInterleave &&
6642 std::max(StoresIC, LoadsIC) > SmallIC) {
6643 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)
;
6644 return std::max(StoresIC, LoadsIC);
6645 }
6646
6647 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)
;
6648 return SmallIC;
6649 }
6650
6651 // Interleave if this is a large loop (small loops are already dealt with by
6652 // this point) that could benefit from interleaving.
6653 bool HasReductions = (Legal->getReductionVars()->size() > 0);
6654 if (TTI.enableAggressiveInterleaving(HasReductions)) {
6655 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)
;
6656 return IC;
6657 }
6658
6659 DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n"
; } } while (false)
;
6660 return 1;
6661}
6662
6663SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
6664LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
6665 // This function calculates the register usage by measuring the highest number
6666 // of values that are alive at a single location. Obviously, this is a very
6667 // rough estimation. We scan the loop in a topological order in order and
6668 // assign a number to each instruction. We use RPO to ensure that defs are
6669 // met before their users. We assume that each instruction that has in-loop
6670 // users starts an interval. We record every time that an in-loop value is
6671 // used, so we have a list of the first and last occurrences of each
6672 // instruction. Next, we transpose this data structure into a multi map that
6673 // holds the list of intervals that *end* at a specific location. This multi
6674 // map allows us to perform a linear search. We scan the instructions linearly
6675 // and record each time that a new interval starts, by placing it in a set.
6676 // If we find this value in the multi-map then we remove it from the set.
6677 // The max register usage is the maximum size of the set.
6678 // We also search for instructions that are defined outside the loop, but are
6679 // used inside the loop. We need this number separately from the max-interval
6680 // usage number because when we unroll, loop-invariant values do not take
6681 // more register.
6682 LoopBlocksDFS DFS(TheLoop);
6683 DFS.perform(LI);
6684
6685 RegisterUsage RU;
6686 RU.NumInstructions = 0;
6687
6688 // Each 'key' in the map opens a new interval. The values
6689 // of the map are the index of the 'last seen' usage of the
6690 // instruction that is the key.
6691 typedef DenseMap<Instruction *, unsigned> IntervalMap;
6692 // Maps instruction to its index.
6693 DenseMap<unsigned, Instruction *> IdxToInstr;
6694 // Marks the end of each interval.
6695 IntervalMap EndPoint;
6696 // Saves the list of instruction indices that are used in the loop.
6697 SmallSet<Instruction *, 8> Ends;
6698 // Saves the list of values that are used in the loop but are
6699 // defined outside the loop, such as arguments and constants.
6700 SmallPtrSet<Value *, 8> LoopInvariants;
6701
6702 unsigned Index = 0;
6703 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
6704 RU.NumInstructions += BB->size();
6705 for (Instruction &I : *BB) {
6706 IdxToInstr[Index++] = &I;
6707
6708 // Save the end location of each USE.
6709 for (Value *U : I.operands()) {
6710 auto *Instr = dyn_cast<Instruction>(U);
6711
6712 // Ignore non-instruction values such as arguments, constants, etc.
6713 if (!Instr)
6714 continue;
6715
6716 // If this instruction is outside the loop then record it and continue.
6717 if (!TheLoop->contains(Instr)) {
6718 LoopInvariants.insert(Instr);
6719 continue;
6720 }
6721
6722 // Overwrite previous end points.
6723 EndPoint[Instr] = Index;
6724 Ends.insert(Instr);
6725 }
6726 }
6727 }
6728
6729 // Saves the list of intervals that end with the index in 'key'.
6730 typedef SmallVector<Instruction *, 2> InstrList;
6731 DenseMap<unsigned, InstrList> TransposeEnds;
6732
6733 // Transpose the EndPoints to a list of values that end at each index.
6734 for (auto &Interval : EndPoint)
6735 TransposeEnds[Interval.second].push_back(Interval.first);
6736
6737 SmallSet<Instruction *, 8> OpenIntervals;
6738
6739 // Get the size of the widest register.
6740 unsigned MaxSafeDepDist = -1U;
6741 if (Legal->getMaxSafeDepDistBytes() != -1U)
6742 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
6743 unsigned WidestRegister =
6744 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
6745 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6746
6747 SmallVector<RegisterUsage, 8> RUs(VFs.size());
6748 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
6749
6750 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)
;
6751
6752 // A lambda that gets the register usage for the given type and VF.
6753 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
6754 if (Ty->isTokenTy())
6755 return 0U;
6756 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
6757 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
6758 };
6759
6760 for (unsigned int i = 0; i < Index; ++i) {
6761 Instruction *I = IdxToInstr[i];
6762
6763 // Remove all of the instructions that end at this location.
6764 InstrList &List = TransposeEnds[i];
6765 for (Instruction *ToRemove : List)
6766 OpenIntervals.erase(ToRemove);
6767
6768 // Ignore instructions that are never used within the loop.
6769 if (!Ends.count(I))
6770 continue;
6771
6772 // Skip ignored values.
6773 if (ValuesToIgnore.count(I))
6774 continue;
6775
6776 // For each VF find the maximum usage of registers.
6777 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
6778 if (VFs[j] == 1) {
6779 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
6780 continue;
6781 }
6782 collectUniformsAndScalars(VFs[j]);
6783 // Count the number of live intervals.
6784 unsigned RegUsage = 0;
6785 for (auto Inst : OpenIntervals) {
6786 // Skip ignored values for VF > 1.
6787 if (VecValuesToIgnore.count(Inst) ||
6788 isScalarAfterVectorization(Inst, VFs[j]))
6789 continue;
6790 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
6791 }
6792 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
6793 }
6794
6795 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)
6796 << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-vectorize")) { dbgs() << "LV(REG): At #" <<
i << " Interval # " << OpenIntervals.size() <<
'\n'; } } while (false)
;
6797
6798 // Add the current instruction to the list of open intervals.
6799 OpenIntervals.insert(I);
6800 }
6801
6802 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
6803 unsigned Invariant = 0;
6804 if (VFs[i] == 1)
6805 Invariant = LoopInvariants.size();
6806 else {
6807 for (auto Inst : LoopInvariants)
6808 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
6809 }
6810
6811 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)
;
6812 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)
;
6813 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)
;
6814 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)
;
6815
6816 RU.LoopInvariantRegs = Invariant;
6817 RU.MaxLocalUsers = MaxUsages[i];
6818 RUs[i] = RU;
6819 }
6820
6821 return RUs;
6822}
6823
6824void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
6825
6826 // If we aren't vectorizing the loop, or if we've already collected the
6827 // instructions to scalarize, there's nothing to do. Collection may already
6828 // have occurred if we have a user-selected VF and are now computing the
6829 // expected cost for interleaving.
6830 if (VF < 2 || InstsToScalarize.count(VF))
6831 return;
6832
6833 // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
6834 // not profitable to scalarize any instructions, the presence of VF in the
6835 // map will indicate that we've analyzed it already.
6836 ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
6837
6838 // Find all the instructions that are scalar with predication in the loop and
6839 // determine if it would be better to not if-convert the blocks they are in.
6840 // If so, we also record the instructions to scalarize.
6841 for (BasicBlock *BB : TheLoop->blocks()) {
6842 if (!Legal->blockNeedsPredication(BB))
6843 continue;
6844 for (Instruction &I : *BB)
6845 if (Legal->isScalarWithPredication(&I)) {
6846 ScalarCostsTy ScalarCosts;
6847 if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
6848 ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
6849
6850 // Remember that BB will remain after vectorization.
6851 PredicatedBBsAfterVectorization.insert(BB);
6852 }
6853 }
6854}
6855
6856int LoopVectorizationCostModel::computePredInstDiscount(
6857 Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
6858 unsigned VF) {
6859
6860 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6861, __PRETTY_FUNCTION__))
6861 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6861, __PRETTY_FUNCTION__))
;
6862
6863 // Initialize the discount to zero, meaning that the scalar version and the
6864 // vector version cost the same.
6865 int Discount = 0;
6866
6867 // Holds instructions to analyze. The instructions we visit are mapped in
6868 // ScalarCosts. Those instructions are the ones that would be scalarized if
6869 // we find that the scalar version costs less.
6870 SmallVector<Instruction *, 8> Worklist;
6871
6872 // Returns true if the given instruction can be scalarized.
6873 auto canBeScalarized = [&](Instruction *I) -> bool {
6874
6875 // We only attempt to scalarize instructions forming a single-use chain
6876 // from the original predicated block that would otherwise be vectorized.
6877 // Although not strictly necessary, we give up on instructions we know will
6878 // already be scalar to avoid traversing chains that are unlikely to be
6879 // beneficial.
6880 if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
6881 isScalarAfterVectorization(I, VF))
6882 return false;
6883
6884 // If the instruction is scalar with predication, it will be analyzed
6885 // separately. We ignore it within the context of PredInst.
6886 if (Legal->isScalarWithPredication(I))
6887 return false;
6888
6889 // If any of the instruction's operands are uniform after vectorization,
6890 // the instruction cannot be scalarized. This prevents, for example, a
6891 // masked load from being scalarized.
6892 //
6893 // We assume we will only emit a value for lane zero of an instruction
6894 // marked uniform after vectorization, rather than VF identical values.
6895 // Thus, if we scalarize an instruction that uses a uniform, we would
6896 // create uses of values corresponding to the lanes we aren't emitting code
6897 // for. This behavior can be changed by allowing getScalarValue to clone
6898 // the lane zero values for uniforms rather than asserting.
6899 for (Use &U : I->operands())
6900 if (auto *J = dyn_cast<Instruction>(U.get()))
6901 if (isUniformAfterVectorization(J, VF))
6902 return false;
6903
6904 // Otherwise, we can scalarize the instruction.
6905 return true;
6906 };
6907
6908 // Returns true if an operand that cannot be scalarized must be extracted
6909 // from a vector. We will account for this scalarization overhead below. Note
6910 // that the non-void predicated instructions are placed in their own blocks,
6911 // and their return values are inserted into vectors. Thus, an extract would
6912 // still be required.
6913 auto needsExtract = [&](Instruction *I) -> bool {
6914 return TheLoop->contains(I) && !isScalarAfterVectorization(I, VF);
6915 };
6916
6917 // Compute the expected cost discount from scalarizing the entire expression
6918 // feeding the predicated instruction. We currently only consider expressions
6919 // that are single-use instruction chains.
6920 Worklist.push_back(PredInst);
6921 while (!Worklist.empty()) {
6922 Instruction *I = Worklist.pop_back_val();
6923
6924 // If we've already analyzed the instruction, there's nothing to do.
6925 if (ScalarCosts.count(I))
6926 continue;
6927
6928 // Compute the cost of the vector instruction. Note that this cost already
6929 // includes the scalarization overhead of the predicated instruction.
6930 unsigned VectorCost = getInstructionCost(I, VF).first;
6931
6932 // Compute the cost of the scalarized instruction. This cost is the cost of
6933 // the instruction as if it wasn't if-converted and instead remained in the
6934 // predicated block. We will scale this cost by block probability after
6935 // computing the scalarization overhead.
6936 unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
6937
6938 // Compute the scalarization overhead of needed insertelement instructions
6939 // and phi nodes.
6940 if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
6941 ScalarCost += TTI.getScalarizationOverhead(ToVectorTy(I->getType(), VF),
6942 true, false);
6943 ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
6944 }
6945
6946 // Compute the scalarization overhead of needed extractelement
6947 // instructions. For each of the instruction's operands, if the operand can
6948 // be scalarized, add it to the worklist; otherwise, account for the
6949 // overhead.
6950 for (Use &U : I->operands())
6951 if (auto *J = dyn_cast<Instruction>(U.get())) {
6952 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6953, __PRETTY_FUNCTION__))
6953 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 6953, __PRETTY_FUNCTION__))
;
6954 if (canBeScalarized(J))
6955 Worklist.push_back(J);
6956 else if (needsExtract(J))
6957 ScalarCost += TTI.getScalarizationOverhead(
6958 ToVectorTy(J->getType(),VF), false, true);
6959 }
6960
6961 // Scale the total scalar cost by block probability.
6962 ScalarCost /= getReciprocalPredBlockProb();
6963
6964 // Compute the discount. A non-negative discount means the vector version
6965 // of the instruction costs more, and scalarizing would be beneficial.
6966 Discount += VectorCost - ScalarCost;
6967 ScalarCosts[I] = ScalarCost;
6968 }
6969
6970 return Discount;
6971}
6972
6973LoopVectorizationCostModel::VectorizationCostTy
6974LoopVectorizationCostModel::expectedCost(unsigned VF) {
6975 VectorizationCostTy Cost;
6976
6977 // Collect Uniform and Scalar instructions after vectorization with VF.
6978 collectUniformsAndScalars(VF);
6979
6980 // Collect the instructions (and their associated costs) that will be more
6981 // profitable to scalarize.
6982 collectInstsToScalarize(VF);
6983
6984 // For each block.
6985 for (BasicBlock *BB : TheLoop->blocks()) {
6986 VectorizationCostTy BlockCost;
6987
6988 // For each instruction in the old loop.
6989 for (Instruction &I : *BB) {
6990 // Skip dbg intrinsics.
6991 if (isa<DbgInfoIntrinsic>(I))
6992 continue;
6993
6994 // Skip ignored values.
6995 if (ValuesToIgnore.count(&I))
6996 continue;
6997
6998 VectorizationCostTy C = getInstructionCost(&I, VF);
6999
7000 // Check if we should override the cost.
7001 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
7002 C.first = ForceTargetInstructionCost;
7003
7004 BlockCost.first += C.first;
7005 BlockCost.second |= C.second;
7006 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)
7007 << 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)
;
7008 }
7009
7010 // If we are vectorizing a predicated block, it will have been
7011 // if-converted. This means that the block's instructions (aside from
7012 // stores and instructions that may divide by zero) will now be
7013 // unconditionally executed. For the scalar case, we may not always execute
7014 // the predicated block. Thus, scale the block's cost by the probability of
7015 // executing it.
7016 if (VF == 1 && Legal->blockNeedsPredication(BB))
7017 BlockCost.first /= getReciprocalPredBlockProb();
7018
7019 Cost.first += BlockCost.first;
7020 Cost.second |= BlockCost.second;
7021 }
7022
7023 return Cost;
7024}
7025
7026/// \brief Gets Address Access SCEV after verifying that the access pattern
7027/// is loop invariant except the induction variable dependence.
7028///
7029/// This SCEV can be sent to the Target in order to estimate the address
7030/// calculation cost.
7031static const SCEV *getAddressAccessSCEV(
7032 Value *Ptr,
7033 LoopVectorizationLegality *Legal,
7034 ScalarEvolution *SE,
7035 const Loop *TheLoop) {
7036 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
7037 if (!Gep)
7038 return nullptr;
7039
7040 // We are looking for a gep with all loop invariant indices except for one
7041 // which should be an induction variable.
7042 unsigned NumOperands = Gep->getNumOperands();
7043 for (unsigned i = 1; i < NumOperands; ++i) {
7044 Value *Opd = Gep->getOperand(i);
7045 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
7046 !Legal->isInductionVariable(Opd))
7047 return nullptr;
7048 }
7049
7050 // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.
7051 return SE->getSCEV(Ptr);
7052}
7053
7054static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
7055 return Legal->hasStride(I->getOperand(0)) ||
7056 Legal->hasStride(I->getOperand(1));
7057}
7058
7059unsigned LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,
7060 unsigned VF) {
7061 Type *ValTy = getMemInstValueType(I);
7062 auto SE = PSE.getSE();
7063
7064 unsigned Alignment = getMemInstAlignment(I);
7065 unsigned AS = getMemInstAddressSpace(I);
7066 Value *Ptr = getPointerOperand(I);
7067 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
7068
7069 // Figure out whether the access is strided and get the stride value
7070 // if it's known in compile time
7071 const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop);
7072
7073 // Get the cost of the scalar memory instruction and address computation.
7074 unsigned Cost = VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
7075
7076 Cost += VF *
7077 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment,
7078 AS, I);
7079
7080 // Get the overhead of the extractelement and insertelement instructions
7081 // we might create due to scalarization.
7082 Cost += getScalarizationOverhead(I, VF, TTI);
7083
7084 // If we have a predicated store, it may not be executed for each vector
7085 // lane. Scale the cost by the probability of executing the predicated
7086 // block.
7087 if (Legal->isScalarWithPredication(I))
7088 Cost /= getReciprocalPredBlockProb();
7089
7090 return Cost;
7091}
7092
7093unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
7094 unsigned VF) {
7095 Type *ValTy = getMemInstValueType(I);
7096 Type *VectorTy = ToVectorTy(ValTy, VF);
7097 unsigned Alignment = getMemInstAlignment(I);
7098 Value *Ptr = getPointerOperand(I);
7099 unsigned AS = getMemInstAddressSpace(I);
7100 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
7101
7102 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7103, __PRETTY_FUNCTION__))
7103 "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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7103, __PRETTY_FUNCTION__))
;
7104 unsigned Cost = 0;
7105 if (Legal->isMaskRequired(I))
7106 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
7107 else
7108 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, I);
7109
7110 bool Reverse = ConsecutiveStride < 0;
7111 if (Reverse)
7112 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7113 return Cost;
7114}
7115
7116unsigned LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I,
7117 unsigned VF) {
7118 LoadInst *LI = cast<LoadInst>(I);
7119 Type *ValTy = LI->getType();
7120 Type *VectorTy = ToVectorTy(ValTy, VF);
7121 unsigned Alignment = LI->getAlignment();
7122 unsigned AS = LI->getPointerAddressSpace();
7123
7124 return TTI.getAddressComputationCost(ValTy) +
7125 TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS) +
7126 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy);
7127}
7128
7129unsigned LoopVectorizationCostModel::getGatherScatterCost(Instruction *I,
7130 unsigned VF) {
7131 Type *ValTy = getMemInstValueType(I);
7132 Type *VectorTy = ToVectorTy(ValTy, VF);
7133 unsigned Alignment = getMemInstAlignment(I);
7134 Value *Ptr = getPointerOperand(I);
7135
7136 return TTI.getAddressComputationCost(VectorTy) +
7137 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
7138 Legal->isMaskRequired(I), Alignment);
7139}
7140
7141unsigned LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I,
7142 unsigned VF) {
7143 Type *ValTy = getMemInstValueType(I);
7144 Type *VectorTy = ToVectorTy(ValTy, VF);
7145 unsigned AS = getMemInstAddressSpace(I);
7146
7147 auto Group = Legal->getInterleavedAccessGroup(I);
7148 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7148, __PRETTY_FUNCTION__))
;
7149
7150 unsigned InterleaveFactor = Group->getFactor();
7151 Type *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor);
7152
7153 // Holds the indices of existing members in an interleaved load group.
7154 // An interleaved store group doesn't need this as it doesn't allow gaps.
7155 SmallVector<unsigned, 4> Indices;
7156 if (isa<LoadInst>(I)) {
7157 for (unsigned i = 0; i < InterleaveFactor; i++)
7158 if (Group->getMember(i))
7159 Indices.push_back(i);
7160 }
7161
7162 // Calculate the cost of the whole interleaved group.
7163 unsigned Cost = TTI.getInterleavedMemoryOpCost(I->getOpcode(), WideVecTy,
7164 Group->getFactor(), Indices,
7165 Group->getAlignment(), AS);
7166
7167 if (Group->isReverse())
7168 Cost += Group->getNumMembers() *
7169 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7170 return Cost;
7171}
7172
7173unsigned LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I,
7174 unsigned VF) {
7175
7176 // Calculate scalar cost only. Vectorization cost should be ready at this
7177 // moment.
7178 if (VF == 1) {
7179 Type *ValTy = getMemInstValueType(I);
7180 unsigned Alignment = getMemInstAlignment(I);
7181 unsigned AS = getMemInstAlignment(I);
7182
7183 return TTI.getAddressComputationCost(ValTy) +
7184 TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, I);
7185 }
7186 return getWideningCost(I, VF);
7187}
7188
7189LoopVectorizationCostModel::VectorizationCostTy
7190LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
7191 // If we know that this instruction will remain uniform, check the cost of
7192 // the scalar version.
7193 if (isUniformAfterVectorization(I, VF))
7194 VF = 1;
7195
7196 if (VF > 1 && isProfitableToScalarize(I, VF))
7197 return VectorizationCostTy(InstsToScalarize[VF][I], false);
7198
7199 Type *VectorTy;
7200 unsigned C = getInstructionCost(I, VF, VectorTy);
7201
7202 bool TypeNotScalarized =
7203 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
7204 return VectorizationCostTy(C, TypeNotScalarized);
7205}
7206
7207void LoopVectorizationCostModel::setCostBasedWideningDecision(unsigned VF) {
7208 if (VF == 1)
7209 return;
7210 for (BasicBlock *BB : TheLoop->blocks()) {
7211 // For each instruction in the old loop.
7212 for (Instruction &I : *BB) {
7213 Value *Ptr = getPointerOperand(&I);
7214 if (!Ptr)
7215 continue;
7216
7217 if (isa<LoadInst>(&I) && Legal->isUniform(Ptr)) {
7218 // Scalar load + broadcast
7219 unsigned Cost = getUniformMemOpCost(&I, VF);
7220 setWideningDecision(&I, VF, CM_Scalarize, Cost);
7221 continue;
7222 }
7223
7224 // We assume that widening is the best solution when possible.
7225 if (Legal->memoryInstructionCanBeWidened(&I, VF)) {
7226 unsigned Cost = getConsecutiveMemOpCost(&I, VF);
7227 setWideningDecision(&I, VF, CM_Widen, Cost);
7228 continue;
7229 }
7230
7231 // Choose between Interleaving, Gather/Scatter or Scalarization.
7232 unsigned InterleaveCost = UINT_MAX(2147483647 *2U +1U);
7233 unsigned NumAccesses = 1;
7234 if (Legal->isAccessInterleaved(&I)) {
7235 auto Group = Legal->getInterleavedAccessGroup(&I);
7236 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~svn301769/lib/Transforms/Vectorize/LoopVectorize.cpp"
, 7236, __PRETTY_FUNCTION__))
;
7237
7238 // Make one decision for the whole group.
7239 if (getWideningDecision(&I, VF) != CM_Unknown)
7240 continue;
7241
7242 NumAccesses = Group->getNumMembers();
7243 InterleaveCost = getInterleaveGroupCost(&I, VF);
7244 }
7245
7246 unsigned GatherScatterCost =
7247 Legal->isLegalGatherOrScatter(&I)
7248 ? getGatherScatterCost(&I, VF) * NumAccesses
7249 : UINT_MAX(2147483647 *2U +1U);
7250
7251 unsigned ScalarizationCost =
7252 getMemInstScalarizationCost(&I, VF) * NumAccesses;
7253
7254 // Choose better solution for the current VF,
7255 // write down this decision and use it during vectorization.
7256 unsigned Cost;
7257 InstWidening Decision;
7258 if (InterleaveCost <= GatherScatterCost &&
7259 InterleaveCost < ScalarizationCost) {
7260 Decision = CM_Interleave;
7261 Cost = InterleaveCost;
7262 } else if (GatherScatterCost < ScalarizationCost) {
7263 Decision = CM_GatherScatter;
7264 Cost = GatherScatterCost;
7265 } else {
7266 Decision = CM_Scalarize;
7267 Cost = ScalarizationCost;
7268 }
7269 // If the instructions belongs to an interleave group, the whole group
7270 // receives the same decision. The whole group receives the cost, but
7271 // the cost will actually be assigned to one instruction.
7272 if (auto Group = Legal->getInterleavedAccessGroup(&I))
7273 setWideningDecision(Group, VF, Decision, Cost);
7274 else
7275 setWideningDecision(&I, VF, Decision, Cost);
7276 }
7277 }
7278}
7279
7280unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
7281 unsigned VF,
7282 Type *&VectorTy) {
7283 Type *RetTy = I->getType();
7284 if (canTruncateToMinimalBitwidth(I, VF))
7285 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
7286 VectorTy = ToVectorTy(RetTy, VF);
7287 auto SE = PSE.getSE();
7288
7289 // TODO: We need to estimate the cost of intrinsic calls.
7290 switch (I->getOpcode()) {
7291 case Instruction::GetElementPtr:
7292 // We mark this instruction as zero-cost because the cost of GEPs in
7293 // vectorized code depends on whether the corresponding memory instruction
7294 // is scalarized or not. Therefore, we handle GEPs with the memory
7295 // instruction cost.
7296 return 0;
7297 case Instruction::Br: {
7298 // In cases of scalarized and predicated instructions, there will be VF
7299 // predicated blocks in the vectorized loop. Each branch around these</